Low Cost Antenna Array Based Drone Tracking Device For Outdoor Environment

LOW COST ANTENNA ARRAY BASED
DRONE TRACKING DEVICE FOR

OUTDOOR ENVIRONMENT
MARCOS TEIXEIRA DE OLIVEIRA
DISSERTAÇÃO DE MESTRADO EM SISTEMAS MECATRÔNICOS

DEPARTAMENTO DE ENGENHARIA MECÂNICA
FACULDADE DE TECNOLOGIA
UNIVERSIDADE DE BRASÍLIA

OUTDOOR ENVIRONMENT
Orientador: Prof. Dr.-Ing. João Paulo C. L. da Costa, PPMEC/UnB

Coorientador: Dr.-Ing Ricardo Kehrle Miranda, PPMEC/UnB
DISSERTAÇÃO DE MESTRADO EM SISTEMAS MECATRÔNICOS
PUBLICAÇÃO PPMEC.DM - XXX/XXXX

BRASÍLIA-DF, 28 DE JANEIRO DE 2019.

OUTDOOR ENVIRONMENT
DISSERTAÇÃO DE MESTRADO ACADÊMICO SUBMETIDA AO DEPARTAMENTO DE

ENGENHARIA MECÂNICA DA FACULDADE DE TECNOLOGIA DA UNIVERSIDADE DE
BRASÍLIA, COMO PARTE DOS REQUISITOS NECESSÁRIOS PARA A OBTENÇÃO DO
GRAU DE MESTRE EM SISTEMAS MECATRÔNICOS.
APROVADA POR:
Prof. Dr.-Ing. João Paulo C. L. da Costa, PPMEC/UnB

Orientador
Prof. Dr. Sebastien Roland Marie Joseph Rondineau, PPMEC/UnB

Examinador interno ao PPMEC
Prof. Dr. Ricardo Zelenovsky, ENE/UnB

Examinador externo ao PPMEC
BRASÍLIA, 28 DE JANEIRO DE 2019.

FICHA CATALOGRÁFICA
Low cost antenna array based drone tracking device for outdoor environments
2019xv, 95p., 201x297 mm
(ENM/FT/UnB, Mestre, Sistemas Mecatrônicos, 2019)
Dissertação de Mestrado - Universidade de Brasília
Faculdade de Tecnologia - Departamento de Engenharia Mecânica
REFERÊNCIA BIBLIOGRÁFICA
MARCOS TEIXEIRA DE OLIVEIRA (2019) Low cost antenna array based drone tracking de-
vice for outdoor environments. Dissertação de Mestrado em Sistemas Mecatrônicos, Publica-
ção XXX/XXXX, Departamento de Engenharia Mecânica, Universidade de Brasília, Brasília,
DF, 95p.
CESSÃO DE DIREITOS
AUTOR: MARCOS TEIXEIRA DE OLIVEIRA

TÍTULO: Low cost antenna array based drone tracking device for outdoor environments.
GRAU: Mestre ANO: 2019
É concedida à Universidade de Brasília permissão para reproduzir cópias desta dissertação de

Mestrado e para emprestar ou vender tais cópias somente para propósitos acadêmicos e cientí-
ficos. Ao autor se reserva outros direitos de publicação e nenhuma parte desta dissertação de
Mestrado pode ser reproduzida sem a autorização por escrito do autor.
____________________________________________________
SHIGS 708 BL. E, Casa 26
Agradecimentos
Gostaria de agradecer primeiramente à minha mãe, Gilda Valente Teixeira Oliveira e

família que desde pequeno sempre me incentivou o caminho de uma boa educação e nunca
mediu esforços para ajudar no que fosse preciso pra conseguir tê-la. Aos meus tios Gilvane
V. T. e Silva e Ivan Silva, que sempre apoiaram e incentivaram a importância do mestrado,
dando conselhos valiosos e o suporte para morar em Brasília que, sem ele, este título não
seria possível. Ao meu padrinho Gilson Valente Teixeira que ao perceber um desinteresse por
minha parte no meio acadêmico, sentou-se comigo colocando toda sua experiência pessoal e
preocupação pedindo para que eu não desistisse.
Gostaria de agradecer também meu co-orientador de Mestrado, Dr.-Ing Ricardo Kehrle
Miranda, que desde o início dedicou seu tempo à acompanhar o desenvolvimento de toda a
pesquisa me passando conceitos básicos por ser uma área nova pra mim, além dos incessantes
falhos e bem sucedidos experimentos em dias de muito sol ou até mesmo chuva. Também
ao meu orientador Prof. Dr.-Ing João Paulo Carvalho Lustosa da Costa que sempre com sua
liderança e visão de um trabalho promissor, conseguiu me guiar em busca de uma pesquisa
de excelência através da interdisciplinarização de áreas como foi o meu caso conciliando um
mundo novo pra mim em Processamento de Sinais e Sistemas Embarcados. Ao Dr. Edison
Pignation de Freitas da Universidade Federal do Rio Grande do Sul pela minha indicação ao
grupo de pesquisa do Laboratory of Array Signal Processing (LASP) da UnB. E também não
posso esquecer ainda de pessoas importantes como a Stephany Ribeiro Rodrigues, Arthur
Carvalho dos Santos, Daniel Akira Ando e João Alvim que também colaboraram bastante
na pesquisa de uma forma não dispensável, tanto através de troca de experiências, como em
fazer os trabalhosos experimentos.
A todos, o meu muito obrigado!
i
Dedicatória
Dedico este trabalho a todos engenheiros, técnicos, profissionais e/ou amadores na área
de Telecomunicações. Que possam fazer bom uso e aproveitar os conhecimentos aqui com-
partilhados para desenvolver e aprimorar tecnologias voltadas para o bem e para a segurança
pública em geral.
ii
Abstract
Applications of Direction of Arrival (DoA) techniques have dramatically increased in

various areas ranging from the traditional wireless communication systems and rescue ope-
rations to GNSS systems and drone tracking. Particularly, police forces and security compa-
nies have drawn their attention to drone tracking devices, due to the number of accidents and
incidents involving Unmanned Aerial Vehicles (UAVs). Now, new systems and devices that
provide the safeness of citizens and clients, have grown and gained space on the market. In
order to detect the presence of drones and to track them, there is a variety of highly expensive
solutions in the market. In this way the estimation of a target’s location can be obtained using
off-the-shelf hardware with Direction of Arrival techniques. Consequently some actions can
be taken by the responsible for the security in that place. Previous works in DoA estimation
using antenna arrays have been proposed but with no practical approach.
In this dissertation, it is proposed a low cost antenna array based drone tracking device
for outdoor environments. The proposed solution is divided into hardware and software
parts. The hardware part of the proposed device is based on off-the-shelf components such
as an omni-directional antenna array, a 4 channel Software Defined Radio (SDR) platform
with carrier frequency ranging from 70 MHz to 6 GHz, a FPGA motherboard and a laptop.
The software part includes algorithms for calibration, model order selection (MOS) and DoA
estimation, including specific pre-processing steps to increase the DoA accuracy. The per-
formance of our proposed low cost solution is evaluated in outdoor scenarios. According to
our measurement campaigns, it is shown that, when the array is in the front fire position, i.e.
with a DoA ranging from -60◦ to 60◦ , the maximum and the average DoA errors are 6◦ and
1,6◦ , respectively.
iii
CONTENTS
1 I NTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 P ROBLEM FORMULATION AND DATA MODEL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3 L ITERATURE R EVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1 M ODEL O RDER SELECTION METHODS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1.1 I NFORMATION T HEORETIC C RITERIA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1.2 S TEIN ’ S U NBIASED R ISK E STIMATOR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.3 RADOI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.4 E XPONENTIAL F ITTING T EST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.5 SAMOS AND ESTER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 P RE - PROCESSING OF DATA METHODS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 S PATIAL S MOOTHING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.2 F ORWARD BACKWARD AVERAGING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.3 VANDERMONDE I NVARIANCE T RANSFORMATION . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 D IRECTION OF A RRIVAL METHODS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1 B EAMFORMINGS AND SUBSPACE BASED METHODS . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1.1 D ELAY AND S UM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1.2 C APON . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1.3 M ULTIPLE S IGNAL C LASSIFICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1.4 ESPRIT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4 L OW COST ANTENNA ARRAY BASED DRONE TRACKING DEVICE AND F RAME -

WORK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.1 H ARDWARE ASSEMBLING FOR THE CALIBRATION . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 H ARDWARE ASSEMBLING FOR THE DRONE TRACKING MEASURE -
MENT CAMPAIGN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3 S IGNAL P ROCESSING F RAMEWORK STRUCTURE . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3.1 P HASE D EVIATION C ORRECTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3.2 BAND -R EJECTION F ILTER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3.3 S ELECTION OF SAMPLES AND M ATRIX X . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3.4 MOS, P RE - PROCESSING AND D OA E STIMATION . . . . . . . . . . . . . . . . . . . . . . . . . . 25
iv
5 R ESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1 E XPERIMENTAL S ETUP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2 BAND -R EJECT F ILTER R ESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3 MOS R ESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.4 D OA R ESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6 C ONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
B IBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
A PPENDIX A P ROPOSED E XTRAPOLATION A LGORITHM TO FIND THE EFT

THRESHOLDS FOR EXTREMELY LOW P ROBABILITY OF FALSE A LARM . . . . . . 50
ATTACHMENT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
A P IX H AWK C ONTROLLER S ETUP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

A.1 ACCELEROMETER C ALIBRATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
A.2 C OMPASS C ALIBRATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
A.3 BATTERY M ONITOR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
A.4 S ONAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
A.5 R ADIO C ALIBRATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
A.6 F LY MODES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
A.7 FAILSAFE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
A.8 G EO F ENCE S ETTINGS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
A.9 P OSSIBLE E RRORS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
R EFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Resumo1
Aplicações para técnicas de Direção de Chegada (DoA) têm crescido drasticamente em

várias áreas, desde os tradicionais sistemas de comunicação sem fio e operações de resgate
até os sistemas GNSS e rastreamento de drones. Particularmente, as forças policiais e as em-
presas de segurança têm voltado sua atenção para os dispositivos de rastreamento de drones,
devido ao número de acidentes e incidentes envolvendo estes Veículos Aéreos não Tripu-
lados (VANTs). Agora, novos sistemas e dispositivos que fornecem segurança a cidadãos
e clientes cresceram e ganharam espaço no mercado. Para detectar a presença de drones e
rastreá-los existe uma variedade de soluções altamente caras no mercado. Porém, a estima-
tiva da localização de um alvo pode ser obtida usando hardware barato, comprado facimente
no mercado, e com técnicas de Direção de Chegada. Data esta estimativa, algumas ações
podem ser tomadas pelo responsável pela segurança no local. Trabalhos anteriores na es-
timativa de direção de chegada usando arranjo de antenas foram propostos, mas sem uma
abordagem prática.
Nesta dissertação, propõe-se um dispositivo de rastreamento de drones baseado em ar-
ranjo de antenas de baixo custo para ambientes externos. A solução proposta é dividida em
partes de hardware e software. A parte de hardware do dispositivo proposto é baseada em
componentes fáceis de serem encontrados no mercado, como um arranjo de antena omnidi-
recional, uma plataforma SDR (Rádio Definido por Software) de 4 canais com frequência
de portadora variando de 70 MHz a 6 GHz, uma placa-mãe FPGA e um laptop. A parte
do software inclui algoritmos para calibração, seleção de ordem de modelo (MOS) e esti-
mativa de DoA, incluindo etapas específicas de pré-processamento para aumentar a precisão
dos cálculos para os métodos de DoA. Avaliamos o desempenho de nossa solução de baixo
custo, proposta para ambientes externos, e de acordo com as medições de campo, mostra-se
que, quando o transmissor está na posição frontal, ou seja, com um DoA variando de -60◦ a
60◦ , o máximo e a média dos erros de DoA são 6◦ e 1,6◦ , respectivamente.
1
EN: This abstract is intended to translate the dissertation into Portuguese language BR, as well as present
a translated and summarized version of the contents of this document. In this chapter, we present a summary
translation of the data model used, the state-of-the-art review, the proposed method and the results obtained
with the experiments.
PT: Este capítulo visa criar um resumo extendido desta dissertação, bem como, apresentar o conteúdo deste
documento em português BR. Apresenta-se uma versão traduzida e resumida do modelo de dados utilizado, da
revisão do estado-da-arte, do método proposto e dos resultados obtidos com a realiação dos experimentos.
vi
Formulação do problema e modelo de dados
Supõe-se d fontes de sinais de banda estreita e suficientemente distantes. Esses sinais

possuem frente de onda planar e atingem um arranjo de antenas de recepção com M sensores
omnidirecionais uniformemente e linearmente dispostos.
O espaço ∆ entre duas antenas adjacentes é igual a λ2 , onde λ é o comprimento de onda do
sinal da portadora. Os sinais recebidos no arranjo de antenas podem ser escritos da seguinte
maneira
X = AS + N ∈ CM ×N , (1)
onde A ∈ CM ×d é a matriz de direção e seu i-ésimo vetor de direção é dado por

h iT
a(µi ) = 1 ejµi e2jµi ... ej(M −1)µi ∈ CM ×1 , (2)
onde µi é a frequência espacial que pode ser mapeada na direção da chegada θi da i-ésima
fonte, de acordo com a seguinte expressão: µi = 2π∆λsin θi . S ∈ Cd×N é a matriz de símbolos
com N sendo o número de amostras. N ∈ CM ×N representa a matriz de ruído, cujos ele-
mentos são supostos como Gaussianos Complexos Circularmente Simétricos, idênticamente
e independentemente distribuídos (i.i.d.).
Dado (1) e assumindo que o ruído e o sinal não estão correlacionados, a matriz de cova-
riância pode ser calculada como
RXX = E{XXH } = ARSS AH + RNN , (3)
onde (·)H é o operador Hermitiano E{·} é o operador do valor esperado. Na prática, a matriz
de covariância da amostra é calculada da seguinte forma
XXH
R̂XX = ∈ CM ×M . (4)
N
As técnicas de DoA usadas ao longo deste documento exploram a matriz de covariância

de amostra em (4). A matriz X é pré-processada antes de se calcular a matriz de covariância
de amostra R̂XX .
O objetivo do nosso dispositivo de rastreamento de drones é estimar a componente Di-
reção de Chegada (DoA) θ1 do sinal em linha de visão direta (LoS) de um drone em um
ambiente externo. Supõe-se que não há obstrução na componente LoS. Portanto, presume-se
que a componente LoS tenha a maior potência em comparação com os componentes não
LoS. Matematicamente, pode-se expressá-la como
||a(µ1 )s(µ1 )||F > ||a(µi )s(µi )||F , (5)
para i = 2, ..., d. O operador || · ||F representa a norma Frobenius e d é a ordem do modelo,

ou seja, d significa o número de fontes ou VANTs presentes no espaço escaneado.
Estado da arte para métodos de estimação de MOS, pré-

processamento e DoA
Nesta Seção são apresentados todos os métodos utilizados para construir o dispositivo
proposto para rastreamento de drones baseado em arranjo de antenas de baixo custo para
ambientes externos. Os métodos de Seleção de Ordem Modelo (MOS) são importantes para
definir o número de Veículos Aéreos Não Tripulados (VANTs) presentes no espaço aéreo
escaneado, e é um pré requisito para alguns algoritmos de DoA. Em seguida, para aprimorar
a precisão dos métdos DoA, os algoritmos de pré-processamento são apresentados. Por fim,
os métodos mais comuns no estado da arte para cálculo da DoA são revisados.
Métodos de seleção da ordem do modelo
Um dos pré-requisitos para a operação adequada do dispositivo proposto é uma boa se-
leção de amostras de dados do sinal obtido para serem processados. Uma das maneiras de
realizar essa seleção é calcular a DoA apenas se estiver na presença de um sinal. Em outras
palavras, deve-se calcular a DoA apenas se existir um VANT no espaço escaneado. Além
disso, vários algoritmos de estimativa DoA dependem do conhecimento prévio da ordem do
modelo. Na Tabela 1 as equações dos métodos usados são resumidamente descritos.
Tabela 1: Métodos MOS utilizados
 1
(M−P)N
 QM
λ M−P 
i=P+1 i
AIC [1] AIC(P) = −log PM + p(M, N) (6)
1
i=P+1 λi
 
M−P
 1
(M−P)N
 QM M−P
i=P+1 λi
 p
MDL [2] EDC(P) = −log PM + p(M, N) Nlog(log(N)) (7)
 1 i=P+1 λi 
M−P
 1
(M−P)N
 QM M−P
i=P+1 λi 1

EDC [3] MDL(P) = −log PM + p(M, N)logN (8)
1 2

M−P i=P+1 λi 
2 2 X P
2 2 4σ̂P σ 1
R̂(P ) = (M − P )σ̂P2 2
+ 2σ P + (σ̂P4 − 2σ̂P σ ) +C
N i=1
λi
4σ 2 X X λi − σˆP2
SURE [4] P M P
2σ 2 2σ 2 σ̂P2
X
C= + P (P − 1) − (M − 1) 1−
N i=1 j=P +1 λi − λj N N i=1
λi
(9)
λP +1 ξP
g1 (P ) = PM , g2 (P ) = PM −1 , P = 1, · · · , M − 1 (10)
i=2 λi i=1 ξi
RADOI [5] M −1

α(λP − µP ) 1 X λP − µP
ξP = 1 − , µP = λi , α = (11)
µP M − P i=P +1 µP
RADOI(P) = g1 (P) − g2 (P) (12)
N
1 − a(P + 1, N ) 1 X
λ̂M −P = (P + 1) · · λM −i (13)
1 − a(P + 1, N )P +1 P + 1 i=0
EFT [6]
v s !
u
u1 15 225 180M
a (M, N ) =t − − (14)
2 M2 + 2 (M 2 + 2)2 N (M 2 − 1) (M 2 + 2)
EE (P ) = W(P )↑ − W(P )↓ Φ(P ),

ESTER [7] (15)
Φ(P ) = (W(P )↓ )† (W(P )↑ ),
U(P ) = [W(P )↑ |W(P )↓ ] ∈ (M − 1) × 2P (16)
SAMOS [8] 2P
1 X tb
ES (P ) = λ (17)
P i=P +1 i
Em (6), (7) e (8) M denota o número de sensores, N é o número de amostras e λi é
o i-ésimo autovalor de R̂XX . Note que a única coisa que muda entre estes 3 métodos é a
parte que acompanha função de penalidade p(M, N). Já em (9) tem-se σ̂P2 sendo a variância
de ruído e esta variância é estimada via Random Matrix Theory (RMT) definido em [4].
Enquanto que em RADOI [5], as funções g1 (P ) e g2 (P ) em (10) são definidas empiricamente
e λ contém os autovalores da matriz de correlação. Passando para EFT [6], λ̂M −P em (13) é
a predição do próximo autovalor de ruído. Em (15) e (16), W(P )↑ contém os P primeiros
autovetores esquerdos em colunas, W(P )↓ contém os últimos P autovetores direitos como
colunas, {·}† denota o operador de pseudo-inversa e λtb
i contém os autovalores de U(P ). Em
todos, P denota o número cadidato a ordem do modelo estimada.
Métodos de pré-processamento de dados
Neste trabalho, os métodos de pré-processamento são aplicados à matriz X

transformando-a em uma matriz Z que será usada pelos outros algoritmos desejados. Desta
forma consegue-se uma precisão maior nos resultados dos métodos usados posteriormente,
no nosso caso os métodos para estimação da DoA. A Tabela 2 expressa as equações utilizadas
pelos métodos descritos.
Tabela 2: Métodos de pré-processamento de dados
(M )
Jl = 0Msub ×(l−1) IMsub 0Msub ×(L−l) , l ∈ 1 ≤ l ≤ L,
(18)
SPS [9]
h i
(M ) (M ) (M )
XSS = J1 X J2 X ... JL X ∈ CMsub ×LN ,
(19)
1
FBA [10, 11] XF B = √ [X, J1 X∗ J2 ], (20)
2
 
1

ejµ(θ) − r
M −1 
 ejν(θ) 

v̌(θ) = Tvit (r)v(θ) = . ..

1−r 
 .


e(M −1)ν(θ)
(21)
VIT [12]
2K.sin(θ)
ν = arctan , (22)
1 − K2 + (1 + K2 ).cos(θ)
Tf (K, θ) = Tvit (r)F(θ). (23)
Y = Tf (r, θ)X ∈ CM ×N . (24)

Na Tabela 2, as matrizes J em (18), (19) e (20) são matrizes de seleção, Msub é o nú-
mero de sensores do sub-arranjo, M o número de sensores do arranjo original e L denota
o número de sub-arranjos. Em (21) e (23), Tvit (r) ∈ CM ×M é a matriz definida em [12],
v(θ) contém a estrutura Vandermonde e ν é a nova fase mapeada da fase original µ. Em
(23), F(θ) = diag{v(θ)} é uma matriz diagonal de foco, contendo a estrutura Vandermonde
mapeada. Finalmente em (24), Y é a matriz de sinal recebida já transformada, pela matriz
de transformação Tf (K, θ), e pronta para ser processada.
Métodos de estimação da Ordem de Chegada
Por fim, a direção de chegada de um sinal é utilizada neste trabalho para a estimativa da
localização e rastreamento de um VANT e alguma ação pode ser tomada pelo responsável
local. Há uma variedade de algoritmos e suas derivações, no entanto, este trabalho se con-
centra nos métodos clássicos mais conhecidos de cálculo de DoA, apresentados na Tabela
3.
Tabela 3: Métodos de estimação DoA selecionados no Estado da Arte
w (θ)H Rxx w (θ)

Delay And Sum [13] PDS (θ) = . (25)
w (θ)H w (θ)
1
Capon [14] PCAP (θ) = H
. (26)
w (θ) R−1
xx w (θ)
1
MUSIC [15] PMUSIC (θ) = , (27)
wH (θ) Vn VnH w (θ)
Ψ = J1 U +
s J2 Us (28)
ESPRIT [16] EVD

Ψ = EΦEH , with Φ = diag[φ1 , ..., φd ] (29)
µi = ∠(φi ), i = 1, ..., d (30)
Na Tabela 3, o vetor w(θ) em (25), (26) e (27) varia de acordo com os valores candidatos
de θ. O valor de θ que maximiza a expressão em (25), (26) e (27) é o θ̂1 , sendo que o
modelo de dados supõe uma componente LoS, ou seja, que não existem obstáculos entre o
transmissor e receptor. Portanto, o componente correspondente à maior potência deve ser
o mesmo componente com DoA θ̂1 . Us ∈ CM ×d é subespaço de sinal, que é igual aos d
autovetores correspondentes aos d maiores autovalores, enquanto que Vn ∈ CM ×M −d é o
subespaço ruído, que é igual aos M − d autovetores correspondentes aos M − d autovalores
com menor potência. Em (29) and (30), Φ é a matriz diagonal que contém os autovalores de
Ψ. Todas as freqüências espaciais são calculadas e a freqüência espacial cujo componente
tem maior potência é o µ̂1 que pode ser mapeado para θ̂1 .
Dispositivo proposto
Nesta seção é detalhado o dispositivo de rastreio de drone baseado em arranjo de antenas

de baixo custo, descrevendo as etapas para a calibração de hardware, colocando todos os
quatro canais do SDR em fase e permitindo a estimativa DoA. Posteriormente, é apresentada
a montagem dos componentes de hardware para o dispositivo. Por fim, um framework do
algoritmo de processamento de sinais para estimativa de DoA é proposto.
Montagem do hardware para calibração
Um dos principais problemas em arranjo de antenas é o desequilíbrio de fase entre as

entradas do SDR. Esse problema pode ser causado por imperfeições de hardware, como
diferentes comprimentos de cabos e trilhas, multipercurso, diferentes instantes de partida
nos osciladores ou até mesmo por simples imperfeições na fabricação. O fornecedor de
hardware disponibiliza um software [17] para sincronização do clock do oscilador local. No
entanto, este software não executa a calibração de fase. Este problema pode ser resolvido por
um filtro que multiplica o sinal recebido por um exponencial contendo o erro de fase. Este
erro de fase é obtido medindo a diferença de fase entre as entradas do SDR, através de um
sinal recebido que foi gerado em uma saída TX do SDR. Este sinal passa por um divisor de
potência com quatro canais de saída de mesma potência e chega em fase a todas as entradas
do SDR.
Montagem do hardware para as medições de campo
Após o hardware ter sido calibrado, o próximo passo é montá-lo para realizar as medi-
ções. O arranjo de antenas omnidirecionais contém quatro sensores e é conectado ao hard-
ware calibrado composto pela placa-mãe FPGA e pela placa-filha SDR. Cada antena é dual
band (de 2400 MHz a 2483,5 MHz e de 4900 MHz a 5875 MHz) [18] e possui polarização
linear com ganho de 3,7 dBi. O espaço ∆ entre duas antenas consecutivas é igual a 59, 9
mm, que dá meio comprimento de onda para a frequência utilizada como portadora.
Estrutura do algoritmo proposto
Correção do desvio de fase
Como citado anteriormente, vários problemas podem causar desequilíbrio de fase. Aqui
é apresentada a abordagem matemática para determinar o vetor de compensação de tal forma
que ele corrija o deslocamento de fase.
Primeiro, a fase dos elementos da matriz X é extraída, da seguinte forma

Im{x(i, j)}
φi,j = ∠x(i, j) = arctan , (31)
Re{x(i, j)}
onde x(i, j) é o elemento na posição i, j da matriz medida X. Os operadores ∠, Im {·} e Re

{·} representam o operador de fase, a parte imaginária de um número complexo e a parte
real de um número complexo, respectivamente.
Para calcular a diferença de fase entre dois canais, a seguinte expressão é aplicada
ω(m, :) = Φ(m, :) − Φ(i, :) ∈ C1×N , (32)
onde Φ representa a matriz com as fases dadas por (31), i indica o canal de referência e m
varia de 1 a M . Dado que o vetor ω(m, :) é a m-ésima linha da matriz Ω, no caso m = i,
a i-ésima linha de Ω é preenchida com zeros. Finalmente, como a diferença de fase pode
variar levemente para diferentes amostras na mesma linha de Ω devido ao ruído térmico, a
média aritmética dos elementos de cada linha de Ω é calculada, obtendo o vetor ω̄ ∈ CM ×1
e seu m-ésimo elemento é dado por
N
1 X
ω̄(m) = Ω(m, n). (33)
N n=1
Assim, para compensar a mudança de fase entre dois canais diferentes, o vetor c ∈ CM ×1
é calculado, cujo m-ésimo elemento é dado por
c(m) = e−j ω̄(m) . (34)
Observe que o vetor de compensação c é calculado apenas uma vez durante a inicializa-
ção do sistema. As saídas calibradas do arranjo de antenas são dadas pela seguinte expressão
Xc = diag{c}X, (35)
onde o operador diag{·} transforma seu vetor argumento na diagonal principal de uma matriz
diagonal.
Filtro rejeita-faixa
No modelo de dados, o ruído é suposto como Gaussiano Circularmente Simétrico de

Valor Complexo e i.i.d. O EFT depende dessas propriedades do ruído. Devido aos valores
extremamente baixos da Probabilidade de Falso Alarme Pfa o comportamento do ruído teve
e ser analisado.
Como a Densidade Espectral de Potência (PSD) mostrada na Fig. 1 não é plana, isto
implica em um ruído correlacionado temporalmente. Além disso percebe-se que o hardware
insere um ruído específico em 1 MHz. Portanto, para remover esse ruído específico de
hardware, um filtro rejeita-faixa deve ser aplicado nessa freqüência.
-20
-20
-30 -30
-40 -40
Poweer [dB]
Poweer [dB]
-50 -50
-60 -60
-70
-70
-80
-80
-90
-90
-1.5 -1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1
Frequency (MHz) Frequency (MHz)
(a) 1a Antena. (b) 2a Antena.
-20 -20
-30 -30
-40 -40
Poweer [dB]
Poweer [dB]
-50 -50
-60
-60
-70
-70
-80
-80
-90
-90
-100
-100
-1.5 -1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1
(c) 3a Antena. (d) 4a Antena.
Figure 1: Densidade Espectral de Potência no caso de apenas ruído para cada antena.
Seleção de amostras
Antes de iniciar o processamento com os algoritmos de pré-processamento e estimadores

de DoA, a autenticidade dos dados deve ser verificada. Portanto, se houver um sinal presente,
a matriz X capturada é passado para a seleção de amostragem explicada nesta subseção.
As fases calculadas de cada entrada do SDR é exemplificada na Fig. 2, e foi empirica-
mente observado que o hardware causa desvios nas fases das amostras em instantes de tempo
aleatórios. Portanto, foi proposta uma abordagem para excluir as amostras com esses desvios
de fase para a estimativa DoA.
4
14 ×10
×104
12 9.2
9
10 8.8
Phase (°)
8.6
8 0 1 2
×105
6 Antenna 1
Antenna 2
Antenna 3
4 Antenna 4
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Samples (time) ×106
Figure 2: Fase dos dados recebidos em cada antena. Um exemplo de região de desvio de
fase é apresentado em destaque.
Observe que há desvios significativos que podem degradar o processo de estimativa DoA.
Como mostrado na Fig. 3, tais desvios podem ser melhor visualizados calculando a diferença
de fase na dimensão do tempo de acordo com a seguinte expressão
γ(m, i) = (φ(m, i + 1) − φ(m, i))2 , (36)
onde γ(m, i) é o valor que contém a diferença quadrática de duas amostras de tempo con-
secutivas i e i + 1 do m-ésimo canal e φ(m,i) é o elemento na posição m,i da matriz Φ da
Equação (31).
3
Derivative of phase (rad)
Antenna1
Antenna2
2.5 Antenna3
Antenna4
Threshold via [111]
2
1.5
0.5
0
3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9
Figure 3: Diferença entre amostras consecutivas da Fig. 2.

Ao detectar os picos na Fig. 3, é possível identificar quais amostras devem ser removidas.
Para esta tarefa, a abordagem proposta em [19] é aplicada. A abordagem em [19] retorna a
linha verde com o valor do limite. Portanto, as amostras cuja diferença de fase é maior que
o limite são removidas.
O resultado após as amostras removidas é apresentado na seguinte equação
Xcfs = [Xcf (:, 1 : N1 )|Xcf (:, N2 : N3 )|...|Xcf (:, NT −1 : NT )], (37)
onde Xcfs é a matriz com as amostras selecionadas. Note que NT = N e os valores de Nt ,

para t = 1, ..., T , são encontrados comparando os valores de diferença de fase com o limite
(linha verde) da Fig. 3.
MOS, Pre-processamento and Estimação DoA
O EFT [20, 6] foi adotado como o método de seleção de ordem de modelo. O EFT tem
as propriedades de deflação que permitem encontrar limites adequados em função da proba-
bilidade de falso alarme (Pfa ). Explorando a propriedade de deflação e encontrando limiares
adequados, o EFT foi o único esquema na literatura analisada a estimar d = 1 na presença de
um forte sinal de LOS e d = 0 nas medições com somente de ruído. Vários outros esquemas
foram comparados como o Akaike Information Criterion (AIC) [1], Efficient Detection Cri-
terion (EDC) [3], Minimum Description Length (MDL) [2], Stein’s Unbiased Risk Estimate
(SURE) [4], RADOI [5], ESTimation ERror (ESTER) [7] e Subspace-based Automatic Mo-
del Order Selection (SAMOS) [8]. O M-EFT [20, 6, 21] também foi adequado, mas um Pfa
ainda menor foi necessário para encontrar os limites. O cálculo dos limites da EFT requer
um Pfa extremamente baixo. Tal cálculo demanda um alto poder computacional. Portanto,
no Apêndice A, é proposto um algoritmo de extrapolação para calcular tais limiares. Observe
que nosso algoritmo de extrapolação proposto foi aplicado em [22, 23, 24], embora nenhum
detalhe seja fornecido nestes trabalhos. Observe que o motivo de um Pfa extremamente baixo
pode estar relacionado ao comportamento de ruído colorido.
A fim de melhorar ainda mais a precisão dos métodos de estimativa DoA, os esquemas
de pré-processamento podem ser aplicados antecipadamente. Neste trabalho foram utili-
zados como esquemas de pré-processamento o Spatial Smoothing (SPS) [9, 25], Forward
Backward Averaging (FBA) [10, 11] e Vandermode Invariation Technique (VIT) [12].
Resultados
Nesta seção, o dispositivo de detecção e rastreamento de drones é validado nas medições

de campo em ambientes externos.
Configuração para Experimento
Na Fig. 4, descrevemos o ambiente externo usado para as medições de campo. No lado

direito, colocamos nosso dispositivo de rastreamento de drone proposto como o receptor,
enquanto que do lado esquerdo, o transmissor é colocado. O transmissor é uma plataforma
MIMO SDR 2x2 ASPR4 [26], com 50 MHz a 6,0 GHz, largura de banda de canal de 200
kHz a 56 MHz e potência máxima de 10 dBm em cada porta de saída. A freqüência central
do transmissor e do receptor baseado em arranjo de antenas foi definida em 2,48 GHz.
Figure 4: Vista de topo do ambiente externo usado para as campanhas de medição, incluindo
as posições do transmissor e do receptor baseado em arranjo de antenas.
O transmissor foi configurado para gerar um sinal de mensagem com modulação MSK
para verificar se o dispositivo está funcionando corretamente. Os símbolos transmitidos são
sequências pseudo-aleatórias com 1024 bits, um cabeçalho 0xFFFF e um rodapé 0x0000. O
dispositivo de rastreiamento de drone proposto foi configurado para uma largura de banda
de 1 MHz, frequência de amostragem de 3 MHz, frequência portadora de 2, 48 GHz, taxa de
dados de 500 kpbs e 5120 amostras.
O transmissor é fixado e o receptor gira de +90◦ até -90◦ em passos de 10◦ . Este passos
são feitos manualmente, portanto o DoA estimado pode ter uma variação devido ao manuseio
do operador que controla a rotação, ou até mesmo erro de impressão no transferidor de papel
apresentado na Fig. 5(b).
Resultado do filtro rejeita-faixa
É importante filtrar o sinal recebido para dar um comportamento gaussiano. Nas Figs. 6
e 7, mostramos o histograma para as antenas 1, 2, 3 e 4, com e sem a filtragem. A melhoria
(a) Posições do transmissor e o receptor baseado em arranjo de antenas. (b) Arranjo de antenas montado
mostrando mostrando o transfe-
ridor de papel.
Figure 5: Fotos tiradas durante a campanha de medição no ambiente externo.
com o filtro é notada com o comportamento Gaussiano apresentado na Fig. 7 com os dados
após a filtragem.
Resultados para MOS
Foram feitos dois experimentos para testar os métodos MOS. O primeiro experimento
usou um transmissor 2x2 MIMO SDR plataforma ASPR4 [26] com potência máxima de 10
dBm. Já o segundo experimento tenta ser mais realista simulando com um roteador [27] um
sinal WiFi usado para o VANT se comunicar com o controlador. Outra vantagem do segundo
experimento é o caso de apenas ruído, já que o sinal não é transmitido de forma contínua,
como ocorre usando o SDR do primeiro experimento.
Para facilitar a visualização, dividimos os resultados de cada experimento em duas ima-
gens, o primeiro experimento é apresentada na Fig. 8 e o segunto experimento é apresentado
na Fig. 9, com o comportamento dos métodos MOS. Observe que colocamos duas figuras
contendo a potência e os autovalores, além das figuras com os resultados do MOS. Com
essas duas imagens, podemos ver o comportamento ao longo do tempo e comparar quando,
de fato, um frame contendo sinal foi recebido e quando há apenas ruído.
3500 3500
Measured noise only data Measured noise only data
Gaussian behavior Gaussian behavior
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
(a) 1a Antena com 70 caixas, média = −0.1035 e des- (b) 2a Antena com 70 caixas, média = −0.0494 e des-
vio padrão = 0.7018. vio padrão = 0.6322.
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
-3 -2 -1 0 1 2 3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
(c) 3a Antena com 70 caixas, média = 0.1162 e desvio (d) 4a Antena com 70 caixas, média = 0.0845 e desvio
padrão = 0.7377. padrão = 0.6836.
Figure 6: Histograma do caso de ruído apenas para cada antena.

3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5
(a) 1a Antena com 70 caixas, média = −0.1035 e des- (b) 2a Antena com 70 caixas, média = −0.0494 e des-
vio padrão = 0.7018. vio padrão = 0.6322.
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 -1.5 -1 -0.5 0 0.5 1 1.5
(c) 3a Antena com 70 caixas, média = 0.1162 e desvio (d) 4a Antena com 70 caixas, média = 0.0845 e desvio
padrão = 0.7377. padrão = 0.6836.
Figure 7: Histograma do caso de ruído filtrado apenas para cada antena.

0
-0.5
-1
-1.5
10 20 30 40 50 60
Time (s)
(a) Potência normalizada média (dB) das 4 antenas de dados capturados em ambiente externo
100
10 20 30 40 50 60
Time (s)
(b) Quatro autovalores (dB) de dados capturados em ambiente externo
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(c) EFT [6] usando os limites dados pelo Appendix A
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(d) Radoi com abordagem discriminante
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(e) SAMOS [8]
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(f) Radoi [5]
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(g) EDC [3]
4
3
MOS 2
1
0
10 20 30 40 50 60
Time (s)
(h) Akaike [1]
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(i) MDL [2]
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(j) Modified EFT [21]
Figure 8: Primeiro experimento para calcular MOS usando um ASRP4 [26] como transmis-
sor.
10
0
-10
50 100 150 200 250 300 350 400 450 500
Time (s)
(a) Potência normalizada média (dB) das 4 antenas de dados capturados em ambiente externo.
100
50 100 150 200 250 300 350 400 450 500

Time (s)
(b) Quatro autovalores de dados capturados em ambiente externo.
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(c) EFT [6] usando os limites dados pelo Appendix A.
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(d) Radoi with discriminant approach
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(e) SAMOS [8]
4
3
MOS 2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(f) EDC [3]
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(g) Modified EFT [21]
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(h) Radoi [5]
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(i) Akaike [1]
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(j) MDL [2]
Figure 9: Segundo experimento para calcular MOS usando um roteador [27] como transmis-
sor.
Exceto para o SAMOS [8] e ESTER [7], a matriz Xcf possui 5120 samples. O proces-
samento computacional para Decomposição em Valores Singulares (SVD) é muito alto para
grandes matrizes, portanto foi usado uma matriz Xcf ∈ C4×1000 para SAMOS and ESTER.
Apesar de alguns pequenos erros, o EFT usando os limites propostos do Apêndice A
tem o melhor desempenho baseado na quantidade de fontes. O Radoi com a abordagem
discriminante também se destaca no ponto em que na presença de um sinal reconheceu a
melhor quantidade de quadros. No entanto, como apresentado na Fig. 10(d) do segundo
experimento, no caso de apenas ruído, ele detecta um número errado de fontes.
Resultados para Direção de Chegada
Esta seção mostra o desempenho dos esquemas DoA usando o ASRP4 [26] como trans-
missor. Durante a campanha de medição, a taxa de erro de bit (BER) alcançada foi de 10−4 .
A Fig. 5.7 mostra o DoA calculado ao girar a base do arranjo de de antennas ao longo
do tempo. Observe que nosso dispositivo proposto funciona corretamente para um DoA que
80 Ds
Capon
Music
60 Esprit
Actual Doa
40
20
DoA (°) 0
-20
-40
-60
-80
50 100 150 200 250 300 350 400 450 500 550
Time (samples)
Figure 10: Comparação entre os métodos de estimativa DoA variando de +90◦ até -90◦ com
passos de 10◦ .
varia de -60◦ a 60◦ . O RMSE, dos métodos DoA já com melhorias de pré-processamento, é
dado por (38) e mostrado na Fig. 11. O pior erro DoA do sistema proposto foi de cerca de
6◦ neste intervalo de -60◦ a 60◦ . A equação do RMSE é dada por
v
u Q
u1 X (q) (q)
RMSE(θ1 ) =t (θ1 − θ̂1 )2 (38)
Q q=1
onde q significa o total de Q realizações para cada 10◦ de passo das medições em campo. Em
cada 10◦ , muitos quadros de N = 5120 amostras são adquiridos para cada canal, nos dando a
matriz X ∈ C4×5120 .
A evolução dos resultados é apresentada nas Tabelas 4, 5, 6, 7, 8 e 9, as quais apresentam
os resultados melhorando com o uso dos esquemas de pré-processamento.
Tabela 4: RMSE para os métodos DoA sem pré-processamento, usando as medidas da Fig.
10
Algoritmo DS CAPON MUSIC ESPRIT

RMSE 2.3388◦ 1.6696◦ 3.9923◦ 2.8080◦
Viriância do RMSE 4.2826◦ 1.0706◦ 11.5493◦ 4.0692◦
Tabela 5: RMSE para os métodos DoA com pré-processamento VIT [12], usando as medi-
ções da Fig. 10

RMSE 2.3388◦ 2.0242◦ 2.8352◦ 2.9575◦
Variância do RMSE 4.2861◦ 1.8770◦ 4.8320◦ 5.2364◦
10
9 Ds
Capon
Music
8 Esprit
RMSE (°)
6
0
-80 -60 -40 -20 0 20 40 60 80
DoA (°)
Figure 11: Comparação do RMSE entre o DoA esperado e o DoA estimado já com pré-
processamento, variando o ângulo entre +90◦ até -90◦ with steps of 10◦ .
Tabela 6: RMSE para os métodos DoA com pré-processamento FBA [10, 11], usando as
medições da Fig. 10

RMSE 2.1205◦ 1.6833◦ 4.0570◦ 2.4022◦
Tabela 7: RMSE para os métodos DoA com pré-processamento SPS [9, 25], usando as

RMSE 1.6484◦ 2.4197◦ 1.9689◦ 1.8716◦
Tabela 8: RMSE para os métodos DoA com pré-processamento FBA e SPS, usando as me-
dições da Fig. 10

RMSE 1.6548◦ 1.6745◦ 1.6813◦ 1.8776◦
Tabela 9: RMSE para os métodos DoA com pré-processamento FBA, SPS e VIT, usando as

RMSE 1.6361◦ 1.6676◦ 1.6475◦ 1.6757◦
Por fim, a evolução do desempenho em médias do RMSE apresentadas nas tabelas 4, 5,

6, 7, 8 e 9 e desenhada na Fig. 12.
Como mostrado na Tabela 10, o custo total da solução proposta para rastreamento de
4.5
Ds
Capon
4 Music
Esprit
3.5
RMSE (°)
3
2.5
1.5
T
al
VI
S
T
S
in
SP
VI
SP
FB
d
ig
an
d
Or
an
S
SP
A
FB
A,
FB
Figure 12: Comparação de desempenho entre os resultados do RMSE apresentados nas
Tabelas 4, 5, 6, 7, 8, 9
drones é US$ 2,222, enquanto as soluções [28] e [29] custam US$ 226.000 e US$ 120.000,
respectivamente. Portanto, nossa solução, já pronta para uso, custa menos de 2 % das solu-
ções comerciais em [28] e [29].
Tabela 10: Tabela de preços do hardware usado na construção do dispositivo proposto de

baixo custo para rastreamento de Drones
Hardware Cost (US$)

AD-FMCOMMS5-EBZ-ND [30] 1080
4 × Dual-Band Antenna 2.4 & 5GHz [18] 11
ZYNQ 7000 Zc702 [31] 999
Power Divider 2.4-6GHz 30Watts RoHS IP67 [32] 109
5 × cable 305mm HPP100 SMA [33] 23
Total 2222
Sumário
1 I NTRODUÇÃO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 F ORMULAÇÃO DO PROBLEMA E MODELO DE DADOS . . . . . . . . . . . . . . . . . . . . . . . 6
3 R EVISÃO BIBLIOGRÁFICA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1 MÉTODOS DE SELEÇÃO DA O RDEM DO M ODELO . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1 C RITÉRIO T EÓRICO DE I NFORMAÇÃO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1.2 S TEIN ’ S U NBIASED R ISK E STIMATOR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.3 RADOI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.4 E XPONENTIAL F ITTING T EST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.5 SAMOS E ESTER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 M ÉTODOS DE PRÉ - PROCESSAMENTO DE DADOS . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 S PATIAL S MOOTHING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.2 F ORWARD BACKWARD AVERAGE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.3 T RANSFORMAÇÃO DE I NVARIÂNCIA DE VANDERMONDE . . . . . . . . . . . . . . . . 14
3.3 M ÉTODOS DE ESTIMAÃO DA O RDEM DE C HEGADA . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1 B EAMFORMING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1.1 ATRASO E S OMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1.2 C APON . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1.3 M ULTIPLE S IGNAL C LASSIFICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4 F RAMEWORK E DISPOSITIVO DE BAIXO CUSTO BASEADO EM ARRAY DE AN -

TENAS PARA RASTREAMENTO DE DRONES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.1 M ONTAGEM DO HARDWARE PARA A CALIBRAÇÃO . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 M ONTAGEM DO HARDWARE PARA CAMPANHA DE RASTREAMENTO
DE DRONE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3 E STRUTURA DO FRAMEWORK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3.1 C ORREÇÃO DE DESVIO DE FASE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3.2 F ILTRO REJEITA BANDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3.3 S ELEÇÃO DE AMOSTRAS E DA MATRIZ X . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3.4 MOS, P RÉ - PROCESSAMENTO E ESTIMAÇÃO DE D OA . . . . . . . . . . . . . . . . . . . . . 25
5 R ESULTADOS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
xxvii
5.1 C ONFIGURAÇÕES PARA OS EXPERIMENTOS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2 R ESULTADOS DO FILTRO REJEITA BANDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3 R ESULTADOS PARA MOS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.4 R ESULTADOS PARA D OA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6 R EFERÊNCIAS B IBLIOGRÁFICAS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
LIST OF FIGURES
1 Densidade Espectral de Potência no caso de apenas ruído para cada antena..... xiv
2 Fase dos dados recebidos em cada antena. Um exemplo de região de desvio
de fase é apresentado em destaque. ........................................................ xv
3 Diferença entre amostras consecutivas da Fig. 2........................................ xv
4 Vista de topo do ambiente externo usado para as campanhas de medição,
incluindo as posições do transmissor e do receptor baseado em arranjo de
antenas. ...........................................................................................xvii
5 Fotos tiradas durante a campanha de medição no ambiente externo. ..............xviii
6 Histograma do caso de ruído apenas para cada antena. ............................... xix
7 Histograma do caso de ruído filtrado apenas para cada antena. ..................... xx
8 Primeiro experimento para calcular MOS usando um ASRP4 [26] como
transmissor. ......................................................................................xxii
9 Segundo experimento para calcular MOS usando um roteador [27] como
transmissor. ......................................................................................xxiii
10 Comparação entre os métodos de estimativa DoA variando de +90◦ até -90◦
com passos de 10◦ . ............................................................................xxiv
11 Comparação do RMSE entre o DoA esperado e o DoA estimado já com pré-
processamento, variando o ângulo entre +90◦ até -90◦ with steps of 10◦ . ........ xxv
12 Comparação de desempenho entre os resultados do RMSE apresentados nas
Tabelas 4, 5, 6, 7, 8, 9 .........................................................................xxvi
2.1 Formulation problem presenting the Uniform Linear Array with with M an-
tennas and drawing impinging signals from an Unmanned Aerial Vehicle....... 6
4.1 Hardware components for building the low cost drone tracking device........... 19
4.2 Radiation pattern for the antenna from Fig. 4.1(d). Image taken from [118] .... 19
4.3 Assembled components for the hardware calibration of four receive chan-
nels. The components are a microprocessor (uProc) and FPGA motherboard,
a SDR, a power divider and cables. ........................................................ 20
4.4 Proposed low cost antenna array based drone tracking device with 4 element
ULA and a 4 x 4 MIMO SDR ............................................................... 21
4.5 Flowchart of proposed framework solution for DoA estimation .................... 22
4.6 Power Density Spectrum of the noise only case for each antenna. ................. 24
xxix
4.7 Phase of the received data in each antenna. An example of phase deviation
region is highlighted in the zoom area..................................................... 25
4.8 Phase difference between consecutive samples from Fig. 4.7. ...................... 26
4.9 Expanded version of box 6 presented in Fig. 4.5 ....................................... 27
5.1 Photos taken during the measurement campaigns of the outdoor scenario. ...... 29
5.2 Top view of the outdoor scenario for the measurement campaigns including
the positions of the transmitter and of the antenna array based receiver .......... 29
5.3 Histogram of the noise only case for each antenna. .................................... 30
5.4 Histogram of the filtered noise only case for each antenna........................... 31
5.5 First experiment to calculate the MOS by schemes from 3.1 using an ASRP4
[26] as transmitter. ............................................................................. 33
5.6 Second experiment to calculate the MOS by schemes from 3.1 using a router
[27] as transmitter. ............................................................................. 34
5.7 Comparison between the DoA estimation schemes by varying the DoA from
+90◦ to -90◦ with steps of 10◦ . ............................................................. 35
5.8 RMSE comparison between the Expected DoA and the DoA estimation com-
bined with pre-processing schemes by varying the angle from +90◦ to -90◦
with steps of 10◦ . ............................................................................... 36
5.9 Comparison of performance between the RMSE results presented in tables
5.1, 5.2, 5.3, 5.4, 5.5, 5.6 ..................................................................... 37
A.1 Log linear extrapolation based on two points given by Pfa s and Thresholds ηP . 51
A.2 Thresholds for extrapolated data using computation with M=4 and N=41. ...... 52
A.1 Selection of the port on Mission Planner software. .................................... 54

A.2 Selection of the frame layout on Mission Planner software. ......................... 55
A.3 Accelerometer Calibration on Mission Planner software. ............................ 55
A.4 Compass Calibration on Mission Planner software. ................................... 56
A.5 Battery selection on Mission Planner software. ......................................... 57
A.6 Sonar selection on Mission Planner software. ........................................... 57
A.7 Radio calibration on Mission Planner software. ........................................ 58
A.8 Flight modes configuration on Mission Planner software. ........................... 61
A.9 Failsafe configuration on Mission Planner software. .................................. 62
A.10 GeoFence configuration on Mission Planner software. ............................... 62
Chapter 1
Introduction
Applications of direction of arrival (DoA) techniques have dramatically increased in var-

ious areas ranging from the traditional wireless communication systems [34, 35] and rescue
operations [36] to GNSS systems [37, 38, 39, 40] and drone tracking in public and private
events. Over the last years, Unmanned Aerial Vehicles (UAVs) have been a major concern
of airspace control bodies and military due to possible terrorist attacks and illegal activi-
ties. In 2015, there were more than nine hundred incidents involving drones and aircrafts
in the United States [41], whereas, in April 2016, an UAV reached an aircraft landing at the
Heathrow airport in London [42]. In Dubai 2016, four drones invaded the airport, one of the
busiest in the world, interrupting the landings and take-offs, causing an estimated loss of one
million dollars [43]. In October 2017 in Canada, the first reported collision of a drone and
a commercial airplane has occurred [44] and, still in 2017, an UAV invaded the Congonhas
airport in São Paulo, Brazil, causing an estimated loss of approximately 1 million Reais due
to the interruption of its services [45]. Most recently, in December 2018 a total of 155 flights
have been canceled and about 10 thousand passengers have been disrupted by drones flying
over one of the UK’s busiest airports, the Gatwick Airport [46]. In addition, [47] and [48]
show cases of drones in football stadiums, threatening the fans safety.
As shown before, these vehicles can, by carelessness, cause millions in losses. Therefore,
recently police forces and security companies have drawn their attention to drone tracking
devices in order to provide the safeness of citizens and clients. In this sense, the development
of low cost devices for drone tracking is fundamental to fit such demands.
In order to detect the presence of drones and to track them, there is a variety of mechan-
ical, optical or antenna array based solutions in the market. For instance, the mechanical
solution in [28] detects a drone within 3 km for targets up to 55 cm in diameter, and clas-
sifies the model of the drone within 1.1 km. The position accuracy (azimuth) in [28] is 1◦ .
In [49], a rechargeable portable drone tracking device can detect and indicate the direction
of a drone in a 360◦ azimuth even with weak Line of Sight (LoS) component. The device
in [49] allows the communication with other devices by using an Application Programming
Interface (API) framework. No technical information and patent about the principles behind
1
the device in [49] and its DoA accuracy are provided. In [29], an antenna array based system
is shown to detect with a 1 km range and with 1◦ accuracy or with a 7 km range and with 3◦
accuracy.
In [50], an application has been proposed for drone detection. According to the develop-
ers, the app has an average range of 106 meters. The system allows the detection of almost
95 % of all types of drones. However, the solution in [50] does not indicate the position or
the direction of the drone.
In [51], a square shaped 16 element antenna array is connected to switches so that a
four channel SDR can select four antennas at each side of the square, allowing a 360◦ DoA
estimation in outdoor environments. Each side of the square performs a ±45◦ azimuth es-
timation. According to the authors, three Yagi antennas were used as sources at specific
points, and maximum DoA error of 5◦ is achieved. No information is provided about the real
distance between the sources and the receive array.
In [52], DoA estimation using an electronically steerable parasitic array radiator (ES-
PAR) with 12 parasitic elements and one active monopole is performed for wireless sensor
network (WSN) applications. The authors calibrated the ESPAR array using an anechoic
chamber. Since the main subject of [52] is the calibration, no measurement outdoor or in-
door campaigns are performed by the authors.
In [53], a four element of quasi-Yagi antenna array system is applied for DoA estimation
using the MUSIC algorithm, whereas the Minimum Description Length (MDL) criterion
is used to estimate the amount of multipath components. Only two measurements are per-
formed for two specific positions showing an error of 1◦ . No information is provided about
the experimental scenario.
In [54], several DoA estimation techniques are compared considering a horizontal uni-
form linear array (ULA) with 12 elements inside an anechoic chamber. The measurements
were conducted varying the DoA from -20◦ to 20◦ in steps of 4◦ . The DoA estimation errors
were smaller than 2◦ . The Min-Norm approach [55] MUSIC [15] although it has a higher
standard deviation.
Finally, in [56], the authors developed system using a five-port reflectometers, that allows
to simultaneously measure the DoA and Time of Arrival (ToA) of coherent and incoherent
signals, connected to seven quasi-Yagi antennas, with one reflectometer for each antenna.
The MUSIC algorithm is applied for the DoA estimation, providing an error of 2◦ for one
source and 0,5◦ for two sources. The measurements were performed in a non-reflective
environment.
In this dissertation, it is proposed a low cost antenna array based drone tracking device
for outdoor environments. To the best of our knowledge, there is no state-of-the-art low cost
off-the-shelf antenna array based device applied to drone tracking. The drone tracking is a
challenging task due to the several possible modulation schemes for the data transmission,
propagation phenomena, such as multipath components, and the possible long operational
2
distances. In the next chapter, the simplified state-of-the-art antenna array based solutions
for DoA estimation in the literature is detailed.
The device proposed is divided into hardware and software parts. The hardware part of
the proposed device is based on off-the-shelf components such as an omni-directional an-
tenna array, a 4 channel Software Defined Radio (SDR) platform, a FPGA motherboard and
a laptop. The software part includes state-of-the-art algorithms for model order selection
(MOS), pre-processing and DoA estimation, including others specific pre-processing and
calibration steps in order to remove random errors and increase the DoA accuracy. The state-
of-the-art algorithm Exponential Fitting Test (EFT) [6] is used for MOS. For pre-processing
schemes, Spatial smoothing (SPS) [9], Forward Backward Averaging (FBA) [10, 11] and
Vandermonde Invariance Transformation (VIT) [12] are used. And finally, the DoA estima-
tion used the Delay and Sum [13], Minimum Variance Distortion-less Response (MVDR) or
Capon’s beamformer [14], MUltiple SIgnal Classification (MUSIC) [15] and Estimation of
Signal Parameters via Rotation Invariance Techniques (ESPRIT) [16] algorithms.
This work also makes a comparative approach between several MOS algorithms in order
to choose the winner, which is used in the proposed device. Additionally presents individual
results of improvement brought by each preprocessing algorithm and its influence on the
accuracy of the DoA estimation.
To build a product, it is necessary to prove the feasibility of it. The frequencies used
for communications by the UAVs in the market are very important in order to select the
frequency range or the array to be used. A survey research, presented in Table 1.1 selected
a variety of popular companies and manufactures of UAVs and its three most highlighted
models for amateur or professional civilian use. Some companies have ’-’ as model name,
that means the company have only one or two models. Note that the most used frequencies
are 2.4 GHz followed by 5 GHz. Therefore the experiments of this dissertation, and the
previous one [57], are focused on 2.4 GHz and 5.8 GHz, respectively.
An additional contribution of this work is to provide the complete configuration of the
PixHawk [58, 59] controller for beginners. This configuration is important for future re-
searchers to be able to continue work using UAVs and can be found in Attachment A.
Building the array of antennas with low-cost hardware is a challenging proposition, since
it deals with difficulties such as mutual coupling, phase calibration and hardware problems.
In the remainder of this work is described each of this problems and the proposed solutions.
This thesis is divided as follows: Chapter 2, the formulation problem and data model is
presented. Next, in Chapter 3, is presented the review of State of the Art for Model Order
Selections Schemes and Direction of Arrival methods that are used in the proposed device.
Then, in Chapter 4, it is proposed a low cost antenna array based drone tracking device
for outdoor environments, including a complete description of the hardware and software,
and the steps involved for assembling, calibration and signal processing. In Section 5, the
proposed solution is validated by means of measurement campaigns in an outdoor scenario
3
presenting the results and methods accuracy. In Section 6, conclusions are drawn.
4
Table 1.1: Table of frequencies (in MHz) used on popular UAVs
Company Model1 (M1) Model2 (M2) Model3 (M3) 433 902-928 2400-2483.5 5150-5850
3D Robotics Solo [60] X8+ [61] IRIS [62] M2 e M3 M2 and M3 M1
AeroEnvironment Quantix [63] - - M1
Autel Robotics EVO [64] X-Star Premium [65] X-Star [65] M2 M1 and M3 M2 and M3
Blade C. 220 BNF Basic [66] Theory XL [67] Torrent 110 [68] M1, M2 and M3
DJI Mavic 2 Pro [69] Phantom 4 Pro [70] Matrice 200 [71] M1, M2 and M3 M1, M2 and M3
EHang GhostDrone [72] Falcon [73] - M1 and M2 M2
Flyability Elios [74] - - M1
GoPro Karma [75, 76] - - M1
Hubsan X4 Air Pro [77] X4 Plus [78] X4 Pro [79] M1, M2 and M3 M1 and M3
Ideafly Storm - 800 [80] Grasshopper F210 [81] Poseidon-480 [82] M1,M2 and M3 M2 and M3
Parrot SA Anafi [83] Bebop 2 [84] Disco [85] M1, M2 and M3 M1, M2 and M3
Powervision PowerEgg [86] PowerEye [87] - M1 and M2
Redcat racing Carbon 210 [88] - - M1 M1
SkyDrones Zangão [89, 90] Pelicano [91, 90] Strix-AG [92, 90] M1 and M2 M3
Syma X5SW [93] X8 pro [94] Z1 [95] M1, M2 and M3
traxxas Aton [96] QR-1 [97] Latrax Alias [98] M1, M2 and M3
Udi RC Discovery 2 plus [99] Peregrine [100] Predator [101] M1, M2 and M3
Walkera Furious 215 [102] Voyager 4 [103] VITUS [104] M1, M2 and M3 M1, M2 and M3
WL toys Q696-A [105] V383 Stuntdrone [106] Q696-D [107] M2 and M3 M1
Yuneec Typhoon 4K [108] Tornado H920 [109] H520 [110] M1,M2 and M3 M1, M2 and M3
5
Chapter 2
Problem formulation and data model
We assume d far filed sources transmitting narrow-band signals. These planar wavefront
signals impinge over a receive antenna array with M omni-directional elements uniformly
and linearly disposed.
Figure 2.1: Formulation problem presenting the Uniform Linear Array with with M antennas
and drawing impinging signals from an Unmanned Aerial Vehicle
The space ∆ between two adjacent antennas is equal to λ2 , where λ is the wavelength of
the carrier signal as presented in Fig. 2.1. The received signals at the antenna array can be
6
written in a matrix fashion as follows
X = AS + N ∈ CM ×N , (2.1)
where A ∈ CM ×d is the steering matrix and its i-th steering vector is given by
h iT
a(µi ) = 1 ejµi e2jµi ... ej(M −1)µi ∈ CM ×1 , (2.2)
where µi is the spatial frequency that can be mapped into the direction of arrival of the i-
th source, θi , by the following expression: µi = 2π∆λsin θi . S ∈ Cd×N is the symbol matrix
with N being the number of snapshots. N ∈ CM ×N stands for the noise matrix, whose ele-
ments are assumed as Complex-Valued Circularly Symmetric Gaussian and identically and
independently distributed (i.i.d.).
Given (2.1) and assuming that the noise and the signal are uncorrelated, the covariance
matrix can be computed by
RXX = E{XXH } = ARSS AH + RNN , (2.3)
where x is one column vector from X, (·)H is the Hermitian operator and E{·} is the expected
value operator. In practice, the sample covariance matrix is calculated as follows:
XXH
R̂XX = ∈ CM ×M . (2.4)
N
The DoA techniques used along this paper exploit the sample covariance matrix in (2.4).
As shown in Chapter 4, the matrix X is pre-processed before the computing of the covariance
matrix R̂XX .
The goal of our proposed drone tracking device is to estimate the Direction of Arrival
(DoA) θ1 of the line of sight (LoS) signal component from a drone in an outdoor scenario.
It is assumed that there is no obstruction of the LoS component. Therefore, the LoS com-
ponent is assumed to have the greatest power in comparison with the Non-LoS components.
Mathematically, it is expressed as
||a(µ1 )s(µ1 )||F > ||a(µi )s(µi )||F , (2.5)
for i = 2, ..., d. The operator || · ||F stands for the Frobenius norm and d is the model order,
in other words d means the number of sources.
7
Chapter 3
Literature Review
In this chapter are presented all the methods used to construct the proposed low cost
antenna array based drones tracking device for outdoor environments. First of all, in Section
3.1 the Model Order Selection (MOS) schemes, which define the number of UAVs present
in the scanned airspace and are a requirement for some DoA schemes, are described. Next,
to provide improved accuracy in DoA schemes, preprocessing algorithms are presented in
Section 3.2. Finally, in Section 3.3 the best-known DoA literature schemes are reviewed.
3.1 Model Order selection methods
One of the prerequisites for proper operation of the proposed device lies in a good se-
lection of the matrix X obtained from the array. In other words, the noise only case needs
to be avoided. One of the ways of this selection is by estimating the order of the model.
Concluding, the device only calculates the Direct of Arrival (DoA) if it is in the presence
of a signal. In addition, several DoA estimation algorithms depend on the knowledge of the
model order.
In the real world we do not know how many signals are arriving in our receiver array,
so several techniques were proposed in the literature in order to estimate this number of
sources. It can be classified as eigenvalues or subspace based, and are usually called model
order selection (MOS) techniques. In this chapter, the used methods are described and the
results of this techniques are drawn in Chapter 5.
3.1.1 Information Theoretic Criteria
Akaike’s Information Criterion (AIC) [1], Minimum Description Length (MDL) [2] and
Efficient Detection Criterion (EDC) [3] have similar expressions with different penalty func-
tions. Therefore they are classified as Information Theoretic Criteria (ITC). Due to the con-
tribution in [111], that give a good performance with the expression for AIC and MDL, they
8
are referred and used as benchmarks. A general expression is used in ITC methods and is
given by

g(P )
J(P ) = −N (M − P )log + p(P, M ), (3.1)
a(P )
ˆ which is obtained minimizing

where P is the candidate value to the estimated model order d,
the J(P ). g(P ) and a(P ) are the geometric and arithmetic mean of P smallest eigenvalues
λi of the ascending sorted main diagonal of a covariance matrix R̂XX defined in (2.4). The
p(P, M ) is a function of the number of independent parameters called penalty function and
each method defines its own.
The MDL log-likelihood criterion is defined by
 1
(M−P)N
 QM M−P
i=P+1 λi 1

MDL(P) = −log PM + p(M, N)logN, (3.2)
 1 i=P+1 λi  2
M−P
where M denotes number of sensors, N is the number of snapshots and λi is the i-th eigen-
value of the R̂XX . p(κ) is the penalty function that change with the proprieties of the matrix
and the preprocessing applied. It is defined as follows
 


 P (2M − P + 1)For real matrices 


 
P (M + P + 1)For real FB - averaging
 

p(M, N ) = . (3.3)


 P (2M − P )For complex matrices 



 
0.5P (2M − P + 1)For complex FB - averaging 
For the AIC, the log-likelihood criterion is given by

 1
(M−P)N
 QM M−P
i=P+1 λi

AIC(P) = −log 1
PM + p(M, N). (3.4)

M−P i=P+1 λi 
Finally, for the EDC, the log-likelihood criterion is given by

 1
(M−P)N
 QM M−P
i=P+1 λi

EDC(P) = −log 1
PM + p(M, N)CN . (3.5)

M−P i=P+1 λi 
where CN can be any function of N such that obey this two proprieties
CN CN
lim = 0, lim = ∞. (3.6)
N →∞ N N →∞ ln(ln(N))
p
We use CN = N log(log(N )) as presented on [7, 112] for which they obtained the best
results.
9
Note that AIC, MDL and EDC are divided in two parts. In the first part, the log-likelihood
criterion function does not change, whilst the second part the penalty function (3.3) changes
with the method. The result of the influence on eigenvalues made by increasing P causes the
first part of the criterion to decrease but at the same time the second part to increase. There-
fore we vary P , in order to find the value that minimizes the AIC(P) and MDL(P). Then, the
estimated number of sources is defined by dˆ = arg min[MDL(P), AIC(P) or EDC(P)].
P
3.1.2 Stein’s Unbiased Risk Estimator
In the more recent Stein’s Unbiased Risk Estimator (SURE) approach [4], P if found in
order to minimize the risk R̂(P ) as follows
2 2 X P
2 2 4σ̂P σ 1
R̂(P ) = (M − P )σ̂P2 2
+ 2σ P + (σ̂P4 − 2σ̂P σ ) +C
N i=1
λi
(3.7)
4σ 2 X X λi − σˆP2
P M P
2σ 2 2σ 2 σ̂P2
X
C= + P (P − 1) − (M − 1) 1− ,
N i=1 j=P +1 λi − λj N N i=1
λi
where σ̂P2 = M 1−P M 2

P
i=P +1 λi and the noise variance σ is estimated via Random Matrix
Theory (RMT) algorithm defined in [4]. In the RMT, to estimate de noise variance, one of
the steps uses a crude estimate of σ 2 with a certain percentile, about 25 and 30-th of smallest
noise corrected eigenvalues, or even the median. Then, the estimated number of sources is
defined by dˆ = arg min R̂(P ).
P
3.1.3 RADOI
According to [5], in the eigenvalue based methods, the method called RADOI performs
better for signals contaminated with colored noise. The main idea is that estimating dˆ is
equivalent to finding how many eigenvalues are associated to each signal and noise sub-
spaces. The empirical functions g1 and g2 are the discriminant functions related to the noise
and signal subspaces respectively, and are given by
λP +1 ξP
g1 (P ) = PM , g2 (P ) = PM −1 , P = 1, · · · , M − 1 (3.8)
i=2 λi i=1 ξi
h i−1
where ξP = 1 − α(λPµ−µ P) 1
PM λP −µP
P
, µP = M −P i=P +1 λi and α = µP
e λ contains the
autocorrelation matrix eigenvalues.
Then, the estimated number of sources dˆ is defined as the first positive value from the
following expression
RADOI(P) = g1 (P) − g2 (P). (3.9)
10
3.1.4 Exponential Fitting Test
The EFT [6] is based on the approximation that the profile of the ordered noise eigenval-
ues have an exponential behavior. The profile a (M, N ) can be expressed as
a (M, N ) =
v s !
(3.10)
u
u1 15 225 180M
t − − .
2 M2 + 2 (M 2 + 2)2 N (M 2 − 1) (M 2 + 2)
Given that dˆ = M − P ∗ , the goal is to vary P such that we find P ∗ that λ̂M −P « λM −P ,
where λ̂M −P and λM −P stands for predicted (M − P )th noise eigenvalue and M − P ac-
tual eigenvalue, respectively. Note that the EFT assumes the smallest eigenvalue as a noise
eigenvalue. Therefore, P varies from 1 to M − 1. Using (3.10), the authors in [6] derived
the following expression

1 − a(P + 1, N )
λ̂M −P = (P + 1) · · σ̂ 2 , (3.11)
1 − a(P + 1, N )P +1
N
2 1 X
σ̂ = λM −i , (3.12)
P + 1 i=0
where σ̂ 2 is the estimated noise power.

In order to improve further the performance of the EFT approach, thresholds coeffi-
cients ηP are computed using noise-only simulated data following Complex-Valued Circu-
larly Symmetric Gaussian and identically and independently distributed (i.i.d.) as indicated
in Chapter 2. Depending on the ηP , there are two hypothesis.

λM −P − λ̂M −P
HP : ≤ ηP (3.13)
λ̂M −P

λM −P − λ̂M −P
H̄P :
> ηP (3.14)
λ̂M −P
where HP : λM −P is a noise eigenvalue and H̄P : λM −P is a signal eigenvalue. In order to

have all ηP depending on the Pfa , it is possible define the Pfa as
Pfa = Pr[d̂ 6= 0|d = 0] (3.15)
where Pr[·] is the probability operator. Note that the ηP thresholds are obtained by Monte
Carlo simulations carried out in the only-noise scenario following the steps in [6] and by
11
choosing the following amount of realizations
10
I= . (3.16)
Pfa
Depending on the noise behaviour and the parameters of the scenario [112, 23], the Pfa
can be extremely low. Therefore, the computational complexity of (3.16) can be prohibitive.
In order to overcome such limitation, it is proposed in Appendix A an extrapolation approach
to compute the thresholds for extremely low values of Pfa .
3.1.5 SAMOS and ESTER
Different from the eigenvalue based, this Subsection presents the subspace based meth-
ods. Subspace based Automatic Model Order Selection (SAMOS) [8] and ESTimation ERror
(ESTER) [7] use the shift invariance propriety and the steering matrix A having the Vander-
monde structure.
ESTER is based on the DoA Estimating algorithm ESPRIT [16], which needs a prior
value of model order d. According to [7], varying the candidate of model order P ∈
{1, · · · , min(M, N)} on algorithm, an erroneous P disturbs the estimation of sinusoids. In
other words, for P 6= d the shift invariance propriety is not valid. Based on this fact, the
authors derive error bounds EE (P ) as follows
EE (P ) = W(P )↑ − W(P )↓ Φ(P ),

(3.17)
Φ(P ) = (W(P )↓ )† (W(P )↑ ),
where W(P )↑ contains the first left P eigenvectors as columns, W(P )↓ contains the last
right P eigenvectors as columns and {·}† denotes the pseudo-inverse operator.
Then the estimated number of sources dˆ from ESTER [7], is defined as the value of P
1
that maximizes the expression (||EE (P )||2 )2
, where || · ||2 stands for the Euclidean vector norm
(or 2-norm).
SAMOS [8] defines a matrix joining the W(P )↑ and W(P )↓ matrices, given expression
U(P ) = [W(P )↑ |W(P )↓ ] ∈ (M − 1) × 2P . According to the Authors, there are tree
hypothesis to be analyzed:
• P > d, the rank of U(P ) is increased by 2;

• P < d, the rank of U(P ) is equal to min(2P,d);
• P = d, the λd+1 , · · · , λ2d singular values of U(P ) are equal to zero.
Based on previous hypothesis the estimated number of order dˆ is the value of P that
12
1
maximizes the expression ES (P )
, where ES (P ) is defined as follows
2P
1 X tb
ES (P ) = λ . (3.18)
P i=P +1 i
3.2 Pre-processing of data methods
In general, the preprocessing methods are used on the matrix X transforming it into a
matrix Z that will be used by the other desired algorithms. Thus, with several concepts
reviewed in this chapter, the preprocessing algorithms allow a better accuracy in the results
of the DoA methods.
3.2.1 Spatial Smoothing
Spatial smoothing (SPS), proposed in [9] and discussed further in [25], has been used in
coherent signals. The spatial smoothing technique introduced by [9] makes it possible to use
DoAs algorithms in presence of arbitrary signal correlation under certain conditions. The
most restrictive aspect os the SPS technique is that it requires an uniformed spaced array.
As our linear antenna array is a 4-element ULA, as well as respecting the above condition,
it can be divided into two sub-arrays with Msub = 3 elements each, {1, · · · , Msub } forming
the first sub-array and {2, · · · , Msub + 1}. It is called Msub the number of sub-array sensors
and M the array number of sensors. Is is applied SPS in X to get a measurement matrix XSS
that has a rank greater than or equal to d. A selection matrix J which selects the elements to
be concatenated, is defined. This approach and XSS is drawn as follows,
(M )
Jl = 0Msub ×(l−1) IMsub 0Msub ×(L−l) , l ∈ 1 ≤ l ≤ L, (3.19)
h i
(M ) (M ) (M )
XSS = J1 X J2 X ... JL X ∈ CMsub ×LN , (3.20)
where L denotes the number of sub-arrays. Note that the number of sensors is reduced from
M to Msub after the spatial smoothing.
3.2.2 Forward Backward Averaging
One of the most commonly used preprocessing schemes, the Forward Backward Averag-
ing (FBA) [10, 11] is used in a wide range of DoA methods. It effectively doubles the data
consequently improving the variances of the estimators. As other methods, FBA has require-
ments to be valid, such as samples be taken in a geometry that is also reversible, arrays must
be centro-symmetric and the properties of the process under consideration be approximately
13
the same independent of the orientation of the time or space axis [10]. With this characteris-
tics, FBA presents the desirable effect of reducing correlation between signals, as occurs in
a multipath environment.
The basic operation of FBA relies on the fact that the matrix A remains the same even if
its elements are reversed and complex conjugated [113]. It is possible to apply the FBA with
this simple equations,
1
XF B = √ [X, J1 X∗ J2 ], (3.21)
2
where J1 ∈ CM ×M and J2 ∈ CN ×N are selection matrices with 1’s on the antidiagonal and
0’s elsewhere.
3.2.3 Vandermonde Invariance Transformation
Different from interpolation methods [114, 115], the Vandermonde Invariance Transfor-
mation (VIT) [12] does not solve the physical array imperfections, it keeps the Vandermonde
structure in a desired direction changing the others, making the uniform Vandermonde re-
sponse of the array into one with a non uniform phase response. It also acts as beamformer
conserving power at a direction of interest and reduces the power from the others.
VIT defines a Tvit (r) ∈ CM ×M and multiplies it by v(θ) =
jµi 2jµi j(M −1)µi T
[1, e , e , · · · , e ] containing a Vandermonde structure. It gives us the
following result
 
1
M −1  ejν(θ)

ejµ(θ) − r
 
v̌(θ) = Tvit (r)v(θ) = . (3.22)
 
1−r .. 

 . 

e(M −1)ν(θ)
where r ∈ C are the complex roots of the polynomials, and can be chosen considering the
noise suppression and the linearity of the output. Note that the original phase µ is mapped
into a new phase ν, that can be expressed as follows

2K.sin(θ)
ν = arctan , (3.23)
1 − K + (1 + K2 ).cos(θ)
2
r+1
where K = r−1
is a phase amplification.
To apply VIT on the received signal, it is built a transformation matrix Tf (K, θ) in order
to transform the received matrix X into a new matrix Y that has the proprieties described
before. This transformation is drawn as follows
Y = Tf (r, θ)X ∈ CM ×N . (3.24)
14
To find Tf , the matrix Tvit is multiplied by a diagonal focusing matrix F(θ) =
diag{v(θ)}, where v(θ) = [1, ejµi , e2jµi , · · · , ej(M −1)µi ] containing a Vandermonde struc-
ture and diag{·} transforms its argument vector into the main diagonal of a diagonal matrix.
Then, the Tf is given by
Tf (K, θ) = Tvit (r)F(θ). (3.25)
Finally, the new data matrix Y is now used by the DoA algorithms obtaining a more
precise DoA estimation with a cost of additional computational load.
3.3 Direction of Arrival methods
The Direction of Arrival of a signal has application in various areas such as radars and
telecommunications. This is further boosted by the recent rise of Unmanned Aerial Vehicles
(UAVs), thereby increasing the concern of airspace control bodies and even governments for
illegal actions and even terrorist attacks. In this way the estimation of a target’s location can
be obtained using the Arrival Direction and some action could be taken. There are a variety
of algorithms and their derivations, however in this work it is focused on the best-known
classical methods of DoA detection, presented in this Section.
3.3.1 Beamformings and subspace based methods
The conventional beamforming methods have the basic idea of "scanning" the array in
a specif direction and measure the output power. When the "scanned" direction is the same
as a DoA of any signal, the output power will reach the maximum. The following equation
address this concept
P (θ) = wH RXX w (3.26)
where {·}H is the Hermitian operator and RXX is the covariance matrix. Most recent beam-
forming methods have been developed by measuring the output power (3.26) with different
forms of the w called spatial filter as it is discussed in this Section.
3.3.1.1 Delay and Sum
For the most simple algorithm Delay and Sum [13] when the scanned direction θ matches
with the real DoA from the source the output power P (θ) will reach a peak or maximum
point.
a (θ)
wDS = p . (3.27)
aH (θ) a (θ)
15
where a(θ) is the steering vector in direction of θ with Vandermonde structure.
By inserting the weight vector (3.27) into (3.26), the output power as a function of angle
of arrival is obtained as follows
a (θ)H RXX a (θ)

PDS (θ) = . (3.28)
a (θ)H a (θ)
This method has a poor resolution. However, to increase its accuracy, it is possible to
add more sensors increasing the amount of storage data. In our case this is not possible since
there is hardware limitations.
3.3.1.2 Capon
The Minimum Variance Distortion-less Response (MVDR) or Capon’s beamformer [14]

achieve the maximum power when the beam with θ matches the direction of a signal. The
weight vector described from Capon is given by,
R−1
XX a(θ)
wCAP = . (3.29)
a(θ) R−1
H
XX a(θ)
Substituting (3.29) into (refeq:principalBeam), the spatial Capon power is given by

1
PCAP (θ) = . (3.30)
a(θ)H R−1
XX a(θ)
According to [113] CAPON works well when there is only one incoming signal present.
However when there is more than one, the array output power contains signal contributions
from the desired and undesired angles. In the particular desired direction, Capon minimizes
the variance of the array output and allows the signal to pass from this direction with no
distortion.
3.3.1.3 Multiple Signal Classification
The last beamformer present in this work is the MUltiple SIgnal Classification (MUSIC)
[15]. It is based on noise subspace, which requires a prior knowledge of number of sources.
The number of sources is used to select the first ordered largest d eigenvalues. Then, the
noise subspace Vn is defined as a matrix formed by the eigenvectors related to the small
ordered eigenvalues λi ∈ {d + 1, · · · , M }.
1
PMUS (θ) = , (3.31)
aH (θ)V H
n Vn a(θ)
As in the previous beamformers, the signals impinging on the array have DoAs given by
the d tallest peaks from the spectrum PMUS (θ).
16
3.3.1.4 ESPRIT
Finally, Estimation of Signal Parameters via Rotation Invariance Techniques (ESPRIT)

[16] is based on two identical subarrays in order to apply the shift invariance staggered in
time. This method is similar to MUSIC at the point that is a subspace based approach, except
to the fact that MUSIC uses the noise subspace and ESPRIT uses the signal subspace.
First, the Singular Value Decomposition (SVD) is applied to the data array X. The
SVD gives us the matrix U ∈ CM ×M of the right singular vectors, that are the eigenvectors
from covariance matrix. Now the signal subspace Us can be found by taking the first d
eigenvectors, sorted according to eigenvalues, from U. In other words, the matrix U can be
rewritten as U = [Us |Un ] with Us ∈ CM ×d .
As the steering matrix A generate the same subspace of Us , there is a nonsingular matrix
T ∈ Cd×d such that A ≈ Us T. The shift invariance propriety give us the assumption
J1 Us TΦ = J2 Us T ⇔ J1 Us Ψ = J2 Us (3.32)
where Ψ = TΦT† ∈ Cd×d is a non singular matrix signal subspace rotating oper-
ator, {·}† being the pseudo-inverse operator, J1 and J2 are the selection matrices and
Φ = diag[φ1 , ..., φd ] is the diagonal matrix that has the eigenvalues representing the esti-
mates of phase factors ejµ containing the DoA information.
As a result, if the signal subspace Us can be estimated, Ψ and its eigenvalues containing
DoAs information can then be found as follows
Ψ = J1 U+
s J2 U s . (3.33)
With Ψ and Φ sharing the same eigenvalues, it is possible to apply the eigenvalue de-
composition (EVD) on Ψ and obtain the φi where 1 ≤ i ≤ M . Finally the DoA is found
computing all the spatial frequencies µ that, if done the array remapping, can be mapped to
θ.
µi = ∠(φi ), i = 1, ..., d (3.34)

λi
θi = arcsin − µi , i = 1, ..., d (3.35)
2π∆
17
Chapter 4
Low cost antenna array based drone

tracking device and Framework
In this Chapter was detailed the proposed low cost antenna array based drone tracking
device. In Section 4.1, the steps for the hardware calibration are described. The calibration
ensure all the four channels of the SDR are in phase, allowing the DoA estimation. In Section
4.2, it was presented the assembling of the components of the hardware of the proposed
drone tracking device. In Section 4.3, a signal processing framework performed for DoA
estimation was proposed.
The hardware components needed to build the low cost drone tracking device are de-
scribed in Table 4.1 and illustrated in the Fig. 4.1.
Table 4.1: List of hardware components used to build the low cost drone tracking device
Item Description
4x4 MIMO platform ADFMCOMMS5 [30] (Fig. 4.1(b)),
with two AD9361 [116] Integrated Circuits (ICs) that contains
Software Defined Radios 2 transmitters and two receivers each, cover 70 MHz
to 6.0 GHz and have a channel bandwidth of 200 kHz
to 56 MHz.
Zynq Zc702 [31] with FPGA Mezzanine Connectors (FMC)
FPGA motherboard allowing scaling and customization with daughter cards
(Fig. 4.1(a)).
Omni-directional, Dual Band 2400–2483.5, 4900–5875
4 antennas MHz, +3.7 dBi of peak gain and linear polarization. [18]
(Fig. 4.1(d) and 4.2).
4 Cables Cable Assembly SMA 305 mm long [33] (Fig. 4.1(c)).
RF Power Splitter, Combiner, Divider with SMA Female
Power division component
4-Way inputs that covers 2.4 - 6.0 GHz
The computer is used to receive all the data from FPGA in
Personal portable computer
order to be processed in the Matlab Software [117]
18
(a) uProc and FPGA motherboard [31] (b) SDR 4x4 MIMO daughterboard [30]
(c) Cable SMA 305mm [33] (d) Dual Band 2.4 & 5GHz (e) 4-Way Splitter 2.4-6GHz
[18] [32]
Figure 4.1: Hardware components for building the low cost drone tracking device
Figure 4.2: Radiation pattern for the antenna from Fig. 4.1(d). Image taken from [118]
4.1 Hardware assembling for the calibration
One of the big problems on practical antenna arrays is phase imbalance between the
inputs of the SDR. This problem may be caused by hardware imperfection such as different
lengths of the cables and tracks, multipath, different time of start on the oscillators or even
by simple imperfections on manufacturing. The hardware vendor provides a software [17]
19
for clock synchronization of the local oscillator. However, this software does not perform
phase calibration. This problem can be solved by a filter that multiplies the signal received
by an exponential of the negatived phase error. This phase error is obtained measuring the
difference of phase between the inputs of an arrived signal that is sent from a TX output,
coming through a power divider with four output signals with the same power and four
cables of equal length. The mathematical approach will be explained in Subsection 4.3.1.
In order to perform the phase calibration, first the hardware components are assembled
according to Fig. 4.3. Note that the SDR transmits a signal from one channel and receives it
in four channels that should be calibrated, such that all receive channels are in phase.
Figure 4.3: Assembled components for the hardware calibration of four receive channels.
The components are a microprocessor (uProc) and FPGA motherboard, a SDR, a power
divider and cables.
As shown in Fig. 4.3, the SDR [30] with two AD9361 [116] is connected to a micropro-
cessor and FPGA motherboard [31], that configures the SDR and transmit the SDR data to
PC through the UTP cable. Note that the cables [33] for calibration should have the same
length. Moreover, a power divider component [32] is included in order to lead the signal to
the four receive channels at the same time, and to reduce the power of the transmitted signal
to avoid damaging.
20
4.2 Hardware assembling for the drone tracking measure-
ment campaign
After the hardware has been calibrated, the next step is to assemble it in order to perform
the measurements. The four elements omni-directional antenna array is connected to the
calibrated hardware composed by the FPGA motherboard and SDR daughterboard according
to Fig. 4.4. Each antenna is dual band (from 2400 MHz to 2483.5 MHz and from 4900 MHz
to 5875 MHz) [18] and has linear polarization with 3.7 dBi of gain. The space ∆ between
two consecutive antennas is equal to 59.9 mm.
Figure 4.4: Proposed low cost antenna array based drone tracking device with 4 element
ULA and a 4 x 4 MIMO SDR
4.3 Signal Processing Framework structure
This section is divided in three subsections. Subsection 4.3.1 explains the mathematical
approach of the hardware calibration of Section 4.1. Then, Subsection 4.3.2 draws the impor-
tance of the filtering. In Subsection 4.3.3 a sample selection approach for DoA estimation by
automatic phase deviation detection is proposed, whereas in Subsection 4.3.4 it is presented
the DoA estimation framework exploiting pre-processing techniques, model order selection
schemes and DoA estimation approaches. Fig. 4.5 depicts the flowchart of proposed signal
21
processing framework solution for DoA estimation.
Figure 4.5: Flowchart of proposed framework solution for DoA estimation
As shown in Box 2 of Fig. 4.5, the phase deviation correction proposed on Subsection
4.3.1 returns a matrix Xc that is used in the filter box from Subsection 4.3.2, in order to get
an improved data. The returned Xcf is now used in sample selection scheme of Subsection
4.3.3, such as it returns Xcfs that is used for pre-processing schemes and DoA estimation in
Subsection 4.3.4.
4.3.1 Phase Deviation Correction
As explained in Section 4.1, several problems may cause phase imbalance. Here it is
presented the mathematical approach in order to find the compensation vector such that it
corrects the phase shift.
First, the phase of the elements of the matrix X is extracted by defining

Im{x(i, j)}
φi,j = ∠x(i, j) = arctan , (4.1)
Re{x(i, j)}
where x(i, j) is the element in position i, j of the measured matrix X. The operators ∠, Im{·}
and Re{·} stand for the phase operator, the imaginary part of a complex number and the real
part of a complex number, respectively.
In order to compute the phase shift between two channels, the following expression is
applied
ω(m, :) = Φ(m, :) − Φ(i, :) ∈ C1×N , (4.2)
where Φ stands for the matrix with the phases given by (4.1), i indicates the reference chan-
nel and m varies from 1 to M . Since the vector ω(m, :) is the m-th row of matrix Ω, in case
22
m = i, the i-th row of Ω is filled with zeros. Finally, since the phase difference may slightly
vary for different samples in the same row of Ω due to the thermal noise, the arithmetic mean
of the elements of each row of Ω is computed, obtaining the vector ω̄ ∈ CM ×1 and its m-th
element is given by
N
1 X
ω̄(m) = Ω(m, n). (4.3)
N n=1
Hence, in order to compensate the phase shift between two different channels, the vector
c ∈ CM ×1 is computed, whose m-th element is given by
c(m) = e−j ω̄(m) . (4.4)
Note that the compensation vector c is computed only once for the system initialization.
The calibrated outputs of the antenna array are given by the following expression
Xc = diag{c}X, (4.5)
where the operator diag{·} transforms its argument vector into the main diagonal of a diag-
onal matrix.
4.3.2 Band-Rejection Filter
In Chapter 2, the noise is assumed to be Complex-Valued Circularly Symmetric Gaussian

and identically and independently distributed (i.i.d.). The EFT relies on these properties
of the noise. Due to extremely low values of the Pfa presented in Appendix A, the noise
behavior is analyzed.
According to Fig. 4.6, the Power Spectrum Density (PSD) is not flat, implying that the
noise is time correlated. Note that the hardware inserts a specific noise at 1 MHz. Therefore,
and according to third box of Flowchart 4.5 to remove this specific noise a band-rejection
filter should be applied at that frequency.
4.3.3 Selection of samples and Matrix X
In order to avoid the unnecessary computational load, when there is no signal present
the framework starts a new capture without execute the pre-processing and DoA algorithms.
This selection is made by checking the Model Order given by the MOS schemes as presented
in Fig. 4.5. Before starting the processing by the pre-processing and DoA algorithms, the
authenticity of the data must also be checked. Therefore, if exists a signal present, the
captured matrix X is passed to the sampling selection that is explained in this subsection.
As exemplified Φ in Fig. 4.7, it is empirically observed that the hardware causes phase
23
-20
-20
-30 -30
-40 -40
Poweer [dB]
Poweer [dB]
-50 -50
-60 -60
-70
-70
-80
-80
-90
-90
-1.5 -1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1
(a) 1st Antenna. (b) 2st Antenna.
-20 -20
-30 -30
-40 -40
Poweer [dB]
Poweer [dB]
-50 -50
-60
-60
-70
-70
-80
-80
-90
-90
-100
-100
-1.5 -1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1
(c) 3st Antenna. (d) 4st Antenna.
Figure 4.6: Power Density Spectrum of the noise only case for each antenna.
deviations on the samples in random time instants. Therefore, an approach to select the
samples with phase deviations for the DoA estimation is proposed.
Note that the phase compensation proposed in Subsection 4.3.1 has been applied on the
samples, whose phases are depicted in Fig. 4.7. Furthermore, note that there are significant
deviations that can degrade the DoA estimation process. As shown in Fig. 4.8, such ripples
can be better visualized by computing the phase difference in the time dimension according
to the following expression
γ(m, i) = (φ(m, i + 1) − φ(m, i))2 , (4.6)
where γ(m, i) is the value containing the quadratic difference of two consecutive time sam-
ples i and i + 1 of the m-th channel, and φ(m,i) is the element in position m,i of the matrix
Φ from (4.1).
By detecting the peaks in Fig. 4.8, it is possible identify which samples should be re-
moved. For this task is used the approach proposed in [19], which returns the value of the
threshold. Therefore, the samples whose phase difference are greater than the threshold are
24
×104
14
×104
12 9.2
9
10 8.8
Phase (°)
8.6
8 0 1 2
×105
6 Antenna 1
Antenna 2
Antenna 3
4 Antenna 4
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Figure 4.7: Phase of the received data in each antenna. An example of phase deviation region
is highlighted in the zoom area.
removed. The result after the samples removed is presented in the following equation
Xcfs = [Xcf (:, 1 : N1 )|Xcf (:, N2 : N3 )|...|Xcf (:, NT −1 : NT )], (4.7)
where Xcfs is the matrix with the selected samples. Note that NT = N and the values of Nt ,
for t = 1, ..., T , are found by comparing the phase difference values with the threshold in
Fig. 4.8.
4.3.4 MOS, Pre-processing and DoA Estimation
According to box 1 Fig. 4.5, the matrix Xcfs given in Subsection 4.3.3 is used now to
improve the DoA estimation with pre-processing schemes. There are several DoA estimation
schemes in the literature as presented in Section 3.3, such as beamforming approaches and
subspace based approaches. In this framework, all the DoA methods presented in the Section
3.3 are adopted, since the redundancy brings confidence and after the pre-processing schemes
the DoA methods have similar performance, as presented in Section 5.4.
The DoA estimation schemes assume that the model order d is known. In practice, model
order selection techniques should be applied to estimate the model order d as depicted in Fig.
25
3
Derivative of phase (rad) Antenna1
Antenna2
2.5 Antenna3
Antenna4
Threshold via [111]
2
1.5
0.5
0
3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9
Figure 4.8: Phase difference between consecutive samples from Fig. 4.7.
4.5.
In the box 4 from Fig. 4.5, the Exponential Fitting Test (EFT) [20, 6] was adopted as
the model order selection scheme. The EFT has the deflation properties that allows to find
suitable thresholds as a function of the Probability of False Alarm (Pfa ). By exploiting the
deflation property and by finding suitable thresholds, the EFT has been the only scheme in
the literature to estimate d = 1 in the presence of a strong LOS signal and d = 0, in the
noise-only measurements. Several schemes were compared in the literature such as Akaike
Information Criterion (AIC) [1], Efficient Detection Criterion (EDC) [3], Minimum Descrip-
tion Length (MDL) [2], Stein’s unbiased risk estimate (SURE) [4], RADOI [5], ESTimation
ERror (ESTER) [7] and Subspace-based Automatic Model Order Selection (SAMOS) [8].
The M-EFT [20, 6, 21] has been also suitable, but a smaller Pfa was required to find the
thresholds. The calculus of the thresholds of the EFT requires an extremely low Pfa . Such
calculus is computationally prohibitive. Therefore, in Appendix A it is proposed an extrapo-
lation algorithm to compute such thresholds. Note that our proposed extrapolation algorithm
has been applied in [22, 23, 24], although no details are provided in these works. Note that
the reason for extremely low Pfa may be related to the colored noise behavior presented in
Subsection 4.3.2.
In order to further improve the accuracy of DoA estimation, pre-processing schemes can
be applied beforehand. In this work the Spatial Smoothing (SPS) [9, 25], Forward Backward
Averaging (FBA) [10, 11] and Vandermode Invariation Technique (VIT) [12] were used as
26
pre-processing schemes. The box 6 from Fig. 4.5 is now enlarged in Fig. 4.9 in order to
best describe the pre-processing step. After that, a matrix Z is returned and used by the DoA
methods summarized in Section 3.3.
Figure 4.9: Expanded version of box 6 presented in Fig. 4.5
27
Chapter 5
Results
In this Chapter, the proposed drone tracking device with measurement campaigns in an
outdoor scenario, is validated. In Section 5.1, the setup for the measurement campaign is
described, while, in Section 5.2, 5.3, and 5.4, are presented the results for the filter, MOS
and DoA respectively.
5.1 Experimental Setup
In Figure 5.1(a), it is depicted the outdoor scenario used for the measurement campaigns.
On the right side, was placed the drone tracking device proposed in Section 4 as the receiver,
while, on the left side, the transmitter is placed. The transmitter, that simulates an UAV, is a
2x2 MIMO SDR platform ASPR4 [26], with 50 MHz to 6.0 GHz, a channel bandwidth of
200 kHz to 56 MHz and max power of 10 dBm at each output port. The center frequency of
both transmitter and antenna array based receiver was set at 2.48 GHz.
As shown in Figure 5.2, the distance between the transmitter and the receiver is 48 m.
Both transmitter and receiver are on tripods 115 cm above the ground. Note that the red “X”
in Figure 5.2 is the location from where the photo in Figure 5.1(a) has been taken.
The transmitter was setup using a MSK message signal to verify that the device functions
properly. The transmitted symbols are pseudo random sequences with 1024 bits, a header
0xFFFF and a footer 0x0000. The proposed drone tracking device was setup to a 1 MHz of
bandwidth, 3 MHz sampling frequency, 2.48 GHz carrier frequency, 500 kpbs data rate and
5120 samples.
As shown in Figure 5.2, the transmitter is fixed and the receiver rotates from +90◦ to
-90◦ in steps of 10◦ . Note that this step is done manually, so the estimated DoA may be
biased due to inaccuracies of the operator controlling the rotation, or even printing error in
the paper-based protractor presented in Figure 5.1(b).
28
(a) Positions of the transmitter and antenna array based receiver (b) Assembled array with paper-
based protractor
Figure 5.1: Photos taken during the measurement campaigns of the outdoor scenario.
Figure 5.2: Top view of the outdoor scenario for the measurement campaigns including the
positions of the transmitter and of the antenna array based receiver
29
5.2 Band-Reject Filter Results
As presented in the Subsection 4.3.2, it is important to filter the received signal in order
to obtain a Gaussian behavior. In Figs. 5.3 and 5.4, it is depicted the histogram for antenna
1, 2, 3 and 4, with and without the filtering. The improvement with the filter is noted with
the Gaussian behavior presented in Fig. 5.4 with the data after filtering.
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
(a) 1st Antenna with 70 bins, mean = −0.1035 and (b) 2st Antenna with 70 bins, mean = −0.0494 and
Standard deviation = 0.7018. Standard deviation = 0.6322.
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
-3 -2 -1 0 1 2 3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
(c) 3st Antenna with 70 bins, mean = 0.1162 and (d) 4st Antenna with 70 bins, mean = 0.0845 and
Figure 5.3: Histogram of the noise only case for each antenna.
30
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5
(a) 1st Antenna with 70 bins, mean = −0.1035 and (b) 2st Antenna with 70 bins, mean = −0.0494 and
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 -1.5 -1 -0.5 0 0.5 1 1.5
(c) 3st Antenna with 70 bins, mean = 0.1162 and (d) 4st Antenna with 70 bins, mean = 0.0845 and
Figure 5.4: Histogram of the filtered noise only case for each antenna.
5.3 MOS Results
The Model Order Selection (MOS) schemes are very important on High Resolution Sig-
nal Processing as well in our proposed device. As explained in Subsection 4.3.3 and 4.3.4 the
MOS is used for DoA schemes and to select a data frame that would be processed in order
to avoid unnecessary data processing if the signal source is not present. The third block of
Flowchart from Fig. 4.5 give the matrix Xcf which is used by MOS schemes.
Two experiments to test the MOS schemes were done. The first experiment used a trans-
mitter 2x2 MIMO SDR platform ASPR4 [26] with max power of 10 dBm, which transmit a
known signal continuously. In the other hand, the second experiment try to be more realistic
and, as presented in Table 1.1, a signal transmitted by an UAV uses WiFi to communicate
with the controller and can be simulated by any router, was used [27]. Other advantage of
the second experiment is the noise case only, with no sources. These experiments used only
one transmitter, so the model order must be always one when a signal is present.
In order to become easier to see, the results is divided into two images for each exper-
31
iment, the first experiment campaign is presented in Fig. 5.5 and the second experiment is
presented in Fig. 5.6, with the behavior of the methods presented in Section 3.1. Note that
the first two figures contain the power and eigenvalues, in addition to the figures with the
MOS results. With these two images it is possible to see the behavior over time and compare
when a signal is received and when there is only noise.
0
-0.5
-1
-1.5
10 20 30 40 50 60
Time (s)
(a) Average normalized power (dB) of the 4 antennas from captured data on outdoor campaign
100
10 20 30 40 50 60
Time (s)
(b) Four Eigen Values (dB) from captured data on outdoor campaingn
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(c) EFT [6] using thresholds from Appendix A
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(d) Radoi-EDC
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(e) SAMOS [8]
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(f) Radoi [5]
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(g) EDC [3]
32
4
3
MOS 2
1
0
10 20 30 40 50 60
Time (s)
(h) Akaike [1]
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(i) MDL [2]
4
3
MOS
2
1
0
10 20 30 40 50 60
Time (s)
(j) Modified EFT [21]
Figure 5.5: First experiment to calculate the MOS by schemes from 3.1 using an ASRP4
[26] as transmitter.
10
0
-10
50 100 150 200 250 300 350 400 450 500
Time (s)
(a) Average normalized power (dB) of the 4 antennas from captured data on outdoor campaingn
100
50 100 150 200 250 300 350 400 450 500

Time (s)
(b) Four eigen values from captured data on outdoor campaingn
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(c) EFT [6] using thresholds from Appendix A
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(d) Radoi-EDC
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(e) SAMOS [8]
33
4
3
MOS 2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(f) EDC [3]
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(g) Modified EFT [21]
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(h) Radoi [5]
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(i) Akaike [1]
4
3
MOS
2
1
0
50 100 150 200 250 300 350 400 450 500
Time (s)
(j) MDL [2]
Figure 5.6: Second experiment to calculate the MOS by schemes from 3.1 using a router
[27] as transmitter.
Except for SAMOS [8] and ESTER [7], the Xcf has 5120 samples. The computational
processing of Singular Value Decomposition (SVD) is very prohibitive for large matrices,
therefore was used a matrix Xcf ∈ C4×1000 matrix for SAMOS and ESTER.
Despite some minor errors, the EFT using the proposed thresholds from Appendix A has
the best performance based on the fact that the recognized number of sources was one. The
Radoi with discriminant approach also stands out at the point where in presence of a signal it
recognized the best amount of frames. However, as presented in the Fig. 5.6(d) of the second
experiment, in the noisy case only it detects the wrong number of sources.
5.4 DoA Results
This Section shows the performance of the DoA schemes shown in Section 3.3 using the
ASRP4 [26] as transmitter. During the measurement campaign, the achieved Bit Error Rate
(BER) was 10−4 .
34
80 Ds
Capon
Music
60 Esprit
Actual Doa
40
20
DoA (°)
-20
-40
-60
-80
50 100 150 200 250 300 350 400 450 500 550
Time (samples)
Figure 5.7: Comparison between the DoA estimation schemes by varying the DoA from
+90◦ to -90◦ with steps of 10◦ .
Fig. 5.7 shows the DoA calculated by rotating the base array over the time. Note that
our proposed device works properly for a DoA ranging from -60◦ to 60◦ . The RMSE, given
by Eq. 5.1, of DoA methods combined with pre-processing schemes is shown in Fig. 5.8.
The worst algorithm DoA error was about 6◦ on this range of -60◦ to +60◦ . The equation for
RMSE is given by
v
u Q
u1 X (q) (q)
RMSE(θ1 ) =t (θ1 − θ̂1 )2 (5.1)
Q q=1
where q stands for one of a total of Q realizations for each 10◦ step of the measurement
campaign. In each 10◦ step, many frames of N = 5120 snapshots are acquired for each
channel, giving us the matrix X ∈ C4×5120 .
The evolution of the results from the range of -60◦ to +60◦ is presented into Tables
5.1, 5.2, 5.3, 5.4, 5.5 and 5.6, which present the results increasing with the use of the pre-
processing schemes.
35
10
9 Ds
Capon
Music
8 Esprit
7
RMSE (°)
0
-80 -60 -40 -20 0 20 40 60 80
DoA (°)
Figure 5.8: RMSE comparison between the Expected DoA and the DoA estimation com-
bined with pre-processing schemes by varying the angle from +90◦ to -90◦ with steps of
10◦ .
Table 5.1: RMSE for the schemes in Section 3.2 without the pre-processing schemes, using
the measurements from Fig. 5.7
Algorithm DS CAPON MUSIC ESPRIT

RMSE 2.3388◦ 1.6696◦ 3.9923◦ 2.8080◦
Variance of the RMSE 4.2826◦ 1.0706◦ 11.5493◦ 4.0692◦
Table 5.2: RMSE for the schemes in Section 3.2 with VIT [12] pre-processing scheme, using
the measurements from Fig. 5.7

RMSE 2.3388◦ 2.0242◦ 2.8352◦ 2.9575◦
Table 5.3: RMSE for the schemes in Section 3.2 with FBA [10, 11] pre-processing scheme,
using the measurements from Fig. 5.7

RMSE 2.1205◦ 1.6833◦ 4.0570◦ 2.4022◦
36
Table 5.4: RMSE for the schemes in Section 3.2 with SPS [9, 25] pre-processing scheme,

RMSE 1.6484◦ 2.4197◦ 1.9689◦ 1.8716◦
Table 5.5: RMSE for the schemes in Section 3.2 with FBA and SPS pre-processing schemes,

RMSE 1.6548◦ 1.6745◦ 1.6813◦ 1.8776◦
Table 5.6: RMSE for the schemes in Section 3.2 with FBA, SPS and VIT pre-processing
schemes, using the measurements from Fig. 5.7

RMSE 1.6361◦ 1.6676◦ 1.6475◦ 1.6757◦
Finally, the performance evolution result RMSE presented in tables 5.1, 5.2, 5.3, 5.4,
5.5 and 5.6 is now drawn in Fig. 5.9. These results show the importance of preprocessing
schemes, which proves that it is better to use all of them in combination.
4.5
Ds
Capon
4 Music
Esprit
3.5
RMSE (°)
2.5
1.5
T
l
VI
na
S
T
S
A
SP
VI
SP
i
FB
d
ig
an
d
Or
an
S
SP
A
FB
,
A
FB
Figure 5.9: Comparison of performance between the RMSE results presented in tables 5.1,
5.2, 5.3, 5.4, 5.5, 5.6
37
As shown in Table 5.7, the total cost of the proposed drone tracking solution is US$
2.222, while the solutions in [28] and [29] cost US$ 226.000 and US$ 120.000, respectively.
Therefore, our proposed off-the-shelf solution costs less than 2 % of the commercial solu-
tions in [28] and [29].
Table 5.7: Price Table for the Hardware of the Proposed Drone Tracker Device
Hardware Cost (US$)

AD-FMCOMMS5-EBZ-ND [30] 1080
4 × Dual-Band Antenna 2.4 & 5GHz [18] 11
ZYNQ 7000 Zc702 [31] 999
Power Divider 2.4-6GHz 30Watts RoHS IP67 [32] 109
5 × cable 305mm HPP100 SMA [33] 23
Total 2222
38
Chapter 6
Conclusion
Over the last years, Unmanned Aerial Vehicles (UAVs) have been a major concern of
airspace control bodies and military due to possible terrorist attacks and illegal activities.
Therefore, recently police forces and security companies have drawn their attention to drone
tracking devices in order to provide the safeness of citizens and clients. Consequently, appli-
cations of direction of arrival (DoA) techniques have dramatically increased in various areas.
In this sense, the development of low cost devices for drone tracking is fundamental to fit
such demands.
In this thesis, was proposed a low cost antenna array based drone tracking device for
outdoor environments. The proposed solution is divided into hardware and software parts.
The hardware of the proposed device is based on off-the-shelf components such as an omni-
directional antenna array, a power division component, four SMA cables, a 4 channel SDR
platform with 70 MHz to 6.0 GHz, a FPGA motherboard and a laptop. The software part in-
cludes state-of-the-art algorithms for model order selection (MOS), pre-processing and DoA
estimation. The performance of the proposed low cost solution was evaluated in outdoor
scenarios. According to our measurement campaigns, it is shown that, when the transmitter
is in the front fire position, i.e. with a DoA ranging from -60◦ to 60◦ , the maximum and the
average DoA errors are 6◦ and 1,6◦ , respectively. Our proposed off-the-shelf solution costs
less than 2 % of commercial solutions in [28] and [29].
This work also makes a comparative approach between several state-of-the-art MOS al-
gorithms in order to choose the method with best performance, which is used in the proposed
device. Despite some minor errors, the state-of-the-art EFT [6] algorithm, using the proposed
thresholds from Appendix A, had the best performance and was used for MOS. Additionally,
presents individual results of improvement brought by each pre-processing algorithm and its
influence on the accuracy of the DoA estimation. It was proved that the DoA methods have a
better accuracy with all the pre-processing schemes such as FBA [10, 11], SPS [9] and VIT
[12], used in combination with each other.
One of the big problems on practical antenna arrays is phase imbalance between the in-
39
puts of the SDR. The presented calibration process aims to solve this problem. Due to minor
random errors caused by the hardware, was also included others specific pre-processing steps
in order to remove the data disturbed, the colored noise and consequently increase the DoA
accuracy.
The DoA estimation used the Delay and Sum [13], CAPON [14], MUSIC [15] and
ESPRIT [16] algorithms, and their accuracy was measured. The DoA methods basically
had similar performance, however, ESPRIT spends more computational load. Without pre-
processing, MUSIC was the one with the highest variance and the worst measured error.
It was also proved that, it is highly possible to detect the direction of an UAV with an
array of antennas built with low cost hardware and without concern about complex issues
such as mutual coupling or calibration in environments with optimal conditions. It is clear
that some improvements can be made, such as to use a more realistic noise model given the
low P fa , and to use some electronic instrument to allow bias reduction, as shown in Section
5.1. However, for our motivation, UAVs can be detected with good accuracy in the range of
-60 to +60 degrees.
40
Bibliography
[1] H. Akaike, “Information Theory as an Extension of the Maximum Likelihood,” in Proc.

2nd Int. Symp. on Information Theory. Proceeding of International Symposium on
Information Theory, 1973, pp. 267–281.
[2] J. Rissanen, “Modeling by shortest data description,” Automatica, vol. 14, no. 5, pp.
465–471, 1978.
[3] L. C. Zhao, P. R. Krishnaiah, and Z. Bai, “On detection of the number of signals in
presence of white noise,” Journal of multivariate analysis, vol. 20, no. 1, pp. 1–25,
1986.
[4] M. O. Ulfarsson and V. Solo, “Rank selection in noisy PCA with SURE and random
matrix theory,” in International Conference on Acoustics, Speech and Signal Processing
(ICASSP). IEEE, 2008, pp. 3317–3320.
[5] E. Radoi and A. Quinquis, “A new method for estimating the number of harmonic com-
ponents in noise with application in high resolution radar,” EURASIP Journal on Applied
Signal Processing, vol. 2004, pp. 1177–1188, 2004.
[6] J. Grouffaud, P. Larzabal, and H. Clergeot, “Some properties of ordered eigenvalues of a

Wishart matrix: application in detection test and model order selection,” in International
Conference on Acoustics, Speech, and Signal Processing (ICASSP)., vol. 5. IEEE, 1996,
pp. 2463–2466.
[7] R. Badeau, B. David, and G. Richard, “Selecting the modeling order for the ESPRIT high
resolution method: an alternative approach,” in International Conference on Acoustics,
Speech, and Signal Processing, vol. 2. IEEE, 2004, pp. ii–1025.
[8] J.-M. Papy, L. De Lathauwer, and S. Van Huffel, “A shift invariance-based order-
selection technique for exponential data modelling,” IEEE signal processing letters,
vol. 14, no. 7, pp. 473–476, 2007.
[9] J. E. Evans, J. R. Johnson, and D. F. Sun, Application of advanced signal

processing techniques to angle of arrival estimation in ATC navigation and
surveillance systems, ser. Technical report (Lincoln Laboratory). Massachusetts
41
Institute of Technology, Lincoln Laboratory, 1982. [Online]. Available: https:
//books.google.com.br/books?id=lOIjSwAACAAJ
[10] D. A. Linebarger, R. D. DeGroat, and E. M. Dowling, “Efficient direction-finding meth-

ods employing forward/backward averaging,” IEEE Transactions on Signal Processing,
vol. 42, no. 8, pp. 2136–2145, 1994.
[11] S. U. Pillai and B. H. Kwon, “Forward/backward spatial smoothing techniques for co-
herent signal identification,” IEEE Transactions on Acoustics, Speech, and Signal Pro-
cessing, vol. 37, no. 1, pp. 8–15, 1989.
[12] T. R. Kurpjuhn, M. T. Ivrlac, and J. A. Nossek, “Vandermonde Invariance Trans-

formation,” in International Conference on Acoustics, Speech, and Signal Processing
(ICASSP)., vol. 5. IEEE, 2001, pp. 2929–2932.
[13] B. D. Van Veen and K. M. Buckley, “Beamforming: A versatile approach to spatial

filtering,” IEEE ASSP magazine, vol. 5, no. 2, pp. 4–24, 1988.
[14] J. Capon, “High-resolution frequency-wavenumber spectrum analysis,” Proceedings of

the IEEE, vol. 57, no. 8, pp. 1408–1418, 1969.
[15] R. Schmidt, “Multiple emitter location and signal parameter estimation,” Transactions
on Antennas and Propagation, vol. 34, no. 3, pp. 276–280, 1986.
[16] R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters via rotational invari-
ance techniques,” Transactions on Acoustics, speech, and signal processing, vol. 37,
no. 7, pp. 984–995, 1989.
[17] A. Devices. (2016) FMComms5 Plugin Description. [Online]. Available: https:

//wiki.analog.com/resources/tools-software/linux-software/fmcomms5_plugin
[18] “Dual band WiFi, 2.4 & 5GHz, PCB.” [Online]. Available: http://www.te.com/usa-en/
product-2118309-1.html
[19] P. F. C. Lima, R. K. Miranda, J. P. C. L. da Costa, R. Zelenovsky, Y. Yuan, and

G. Del Galdo, “Low complexity blind separation technique to solve the permutation
ambiguity of convolutive speech mixtures,” in 10th International Conference on Signal
Processing and Communication Systems (ICSPCS). IEEE, 2016, pp. 1–10.
[20] A. Quinlan, J.-P. Barbot, P. Larzabal, and M. Haardt, “Model order selection for short
data: An exponential fitting test (EFT),” EURASIP Journal on Applied Signal Process-
ing, vol. 2007, no. 1, pp. 201–201, 2007.
[21] J. P. C. L. da Costa, M. Haardt, F. Romer and G. Del Galdo, “Enhanced model order
estimation using higher-order arrays,” in Conference Record of the Forty-First Asilomar
Conference on Signals, Systems and Computers (ACSSC). IEEE, 2007, pp. 412–416.
42
[22] J. P. C. L. da Costa, M. Haardt, F. Romer, and G. Del Galdo, “Enhanced model order
estimation using higher-order arrays,” in 2007 IEEE Conference Record of the Forty-
First Asilomar Conference on Signals, Systems and Computers (ACSSC). IEEE, 2007,
pp. 412–416.
[23] J. P. C. L. da Costa, F. Roemer, M. Haardt, and R. T. de Sousa, “Multi-dimensional

model order selection,” EURASIP Journal on Advances in Signal Processing, vol. 2011,
no. 1, p. 26, 2011.
[24] J. P. C. L. da Costa, M. Haardt, and F. Romer, “Robust methods based on the HOSVD
for estimating the model order in PARAFAC models,” in 5th IEEE Sensor Array and
Multichannel Signal Processing Workshop (SAM). IEEE, 2008, pp. 510–514.
[25] T.-J. Shan, M. Wax, and T. Kailath, “On spatial smoothing for direction-of-arrival es-
timation of coherent signals,” IEEE Transactions on Acoustics, Speech, and Signal Pro-
cessing, vol. 33, no. 4, pp. 806–811, 1985.
[26] “SDR Starter Kit 2x2 MIMO 50 MHz to 6 GHz PC based.” [Online]. Available:
http://akashkosgi.wixsite.com/agile-solutions/sdr-starter-kit-c1a59
[27] “Elecom WRC-300FEBK IEEE 802.11n/g/b 300Mbps.” [Online]. Available: http:

//www2.elecom.co.jp.e.gj.hp.transer.com/products/WRC-300FEBK.html?
[28] “Robin Radar Systems BV, Elvira R Drone Detection Radar.” [Online]. Available:
https://www.robinradar.com/markets/drone-detection/
[29] “Real-Time RF Drone/UAV and Radar Detection System.” [Online]. Available:

http://www.aaronia.com/products/solutions/Aaronia-Drone-Detection-System/
[30] “Eval board software Defined Radio ADFMCOMMS5-EBZ-ND.” [Online].

Available: http://www.analog.com/en/design-center/evaluation-hardware-and-software/
evaluation-boards-kits/eval-ad-fmcomms5-ebz.html
[31] “ZYNQ-7000 ZC702 Eval KIT.” [Online]. Available: https://www.xilinx.com/

products/boards-and-kits/ek-z7-zc702-g.html
[32] “RF Power Splitter, combiner, divider - IP67, 4 Way, SMA Female, 2.4 -
6.0 GHz.” [Online]. Available: https://www.instockwireless.com/rf_power_splitter_
combiner_divider_pd2578.htm
[33] “Cable assembly SMA long 305 mm.” [Online]. Available: https:
//www.instockwireless.com/
[34] R. K. Miranda, J. P. C. L. da Costa, and F. Antreich, “High accuracy and low complex-
ity adaptive generalized Sidelobe Cancelers for colored noise scenarios,” Digital Signal
Processing, vol. 34, pp. 48–55, 2014.
43
[35] R. K. Miranda, J. P. C. L. da Costa, F. Roemer, L. R. A. X. Menezes, G. Del Galdo,
and A. L. F. de Almeida, “Low complexity performance assessment of a sensor array via
unscented transformation,” Digital Signal Processing, vol. 63, pp. 190–198, 2017.
[36] D. Neudert-Schulz, “A contribution to efficient direction finding using antenna arrays,”

Ph.D. dissertation, Technische Universität Ilmenau, 2017.
[37] M. A. M. Marinho, F. Antreich, and J. P. C. L. da Costa, “Improved array interpolation

for reduced bias in DOA estimation for GNSS,” Proc. of Institute of Navigation (ION)
GNSS+, 2014.
[38] D. V. de Lima, J. P. C. L. da Costa, F. Antreich, R. K. Miranda, and G. Del Galdo, “High

resolution Time-delay estimation via direction of arrival estimation and Khatri-Rao fac-
torization for multipath mitigation,” in 21st International ITG Workshop on Smart An-
tennas (WSA 2017), Berlin, Germany, 2017, pp. 388–395.
[39] D. V. de Lima, J. P. C. L. da Costa, J. P. A. Maranhão, and R. T. de Sousa, “Time-delay

estimation via procrustes estimation and Khatri-Rao factorization for GNSS multipath
mitigation,” in 2017 11th International Conference on Signal Processing and Communi-
cation Systems (ICSPCS). IEEE, 2017, pp. 1–7.
[40] D. V. de Lima, J. P. C. L. da Costa, F. Antreich, R. K. Miranda, and G. Del Galdo,

“Time-Delay estimation via CPD-GEVD applied to tensor-based GNSS arrays with er-
rors,” in 7th International Workshop on Computational Advances in Multi-Sensor Adap-
tive Processing (CAMSAP). IEEE, 2017, pp. 1–5.
[41] F. Karimi. (2015) Hundreds of drones fly dangerously close to manned aircraft.
[Online]. Available: https://edition.cnn.com/2015/12/12/us/drone-aircraft-close-calls/
[42] K. Rawlinson. (2016) Drone hits plane at Heathrow airport, says

pilot. [Online]. Available: https://www.theguardian.com/uk-news/2016/apr/17/
drone-plane-heathrow-airport-british-airways
[43] Z. Alkhalisi. (2016) Dubai deploys a ’drone hunter’ to keep its airport open. [Online].
Available: http://money.cnn.com/2016/11/04/technology/dubai-airport-drone-hunter/
index.html?iid=EL
[44] J. Goglia. (2017) A small drone hits a commercial airliner ... and nothing
happens. [Online]. Available: https://www.forbes.com/sites/johngoglia/2017/10/19/
a-small-drone-hits-a-commercial-airliner-and-nothing-happens/#78ad9a249ea1
[45] G. SP. (2017) Drone sobre Congonhas pode causar pre-

juízo de mais de R$ 1 milhão, estima associação de empre-
sas aéreas. [Online]. Available: https://g1.globo.com/sao-paulo/noticia/
drone-sobre-congonhas-causou-prejuizo-de-mais-de-r-1-milhao-estima-associacao.
ghtml
44
[46] BBC. (2018) Gatwick airport: Drones ground flights. [Online]. Available:
https://www.bbc.com/news/uk-england-sussex-46623754
[47] D. Lisboa. (2015) Drone invade arena do Corinthians. é ilegal em estádios e tem
riscos. [Online]. Available: https://esporte.uol.com.br/futebol/ultimas-noticias/2015/05/
20/drone.htm?cmpid=copiaecola
[48] GloboEsporte.com. (2016) Torcedores colorados depredam

casa após drone da “b" sobrevoar Beira-Rio. [Online].
Available: http://globoesporte.globo.com/rs/futebol/noticia/2016/11/
torcedores-colorados-depredam-casa-apos-drone-da-b-sobrevoar-beira-rio.html
[49] “Dronelabs LLC, DD610AR Stationary Drone Detector.” [Online]. Available:

http://dronedetector.com/stationary-unit/
[50] “Detect, Inc., Drone Watcher.” [Online]. Available: www.dronewatcher.com/
[51] M.-C. Hua, C.-H. Hsu, W.-J. Liao, C.-C. Yao, T.-H. Yeh, and H.-C. Liu, “Direction-
of-arrival estimator using array switching on software defined radio platform,” in In-
ternational Symposium on Antennas and Propagation (APSURSI). IEEE, 2011, pp.
2821–2824.
[52] L. Kulas, “RSS-based DoA estimation using ESPAR antennas and interpolated radi-
ation patterns,” Antennas and Wireless Propagation Letters, vol. 17, no. 1, pp. 25–28,
2018.
[53] A. B. Delai and V. Y. Vu, “Digital solution for inter-vehicle localization system by
means of direction-of-arrival,” in International Symposium on Intelligent Signal Pro-
cessing and Communications (ISPACS). IEEE, 2006, pp. 875–878.
[54] S. Bjorklund and A. Heydarkhan, “High resolution direction of arrival estimation meth-
ods applied to measurements from a digital array antenna,” in Proceedings of the Sensor
Array and Multichannel Signal Processing Workshop. IEEE, 2000, pp. 464–468.
[55] H. Krim and M. Viberg, “Two decades of array signal processing research: the para-
metric approach,” Signal Processing Magazine, vol. 13, no. 4, pp. 67–94, 1996.
[56] A. J. Braga, X. Begaud, B. Huyart and V. Y. Vu, “Direction of arrival and time of ar-
rival measurements using five-port reflectometers and quasi-Yagi antennas,” in 11th In-
ternational Symposium on Antenna Technology and Applied Electromagnetics [ANTEM
2005]. IEEE, 2005, pp. 1–4.
[57] D. A. Ando, Direction of Arrival estimation using Off-the-Shelf directional antennas

array, 2018. [Online]. Available: https://lasp.unb.br/wp-content/uploads/files/research/
Daniel_Akira_relatorio_final.pdf
45
[58] L. Meier, P. Tanskanen, L. Heng, G. H. Lee, F. Fraundorfer, and M. Pollefeys, “Pix-
hawk: A micro aerial vehicle design for autonomous flight using onboard computer
vision,” Autonomous Robots, vol. 33, no. 1-2, pp. 21–39, 2012.
[59] L. Meier, D. Honegger, and M. Pollefeys, “Px4: A node-based multithreaded open

source robotics framework for deeply embedded platforms,” in Robotics and Automation
(ICRA), 2015 IEEE International Conference on. IEEE, 2015, pp. 6235–6240.
[60] 3DR. (2015) SOLO User Manual. [Online]. Available: http://3dr.com/wp-content/

uploads/2017/03/v9_02_25_16.pdf
[61] 3DR. (2014) X8+ Operation Manual. [Online]. Available: http://3dr.com/wp-content/

uploads/2017/03/X8-Operation-Manual-vA.pdf
[62] 3DR. (2014) IRIS Operation Manual. [Online]. Available: http://3dr.com/wp-content/

uploads/2017/03/IRIS-Operation-Manual-v6.pdf
[63] AeroVironment. (2017) Quantix User Manual. [Online]. Available: https://www.

avdroneanalytics.com/Upload/MediaCenter/636675046702100584.pdf
[64] Autel Robotics. (2017) EVO page. [Online]. Available: https://www.autelrobotics.

com/evo/
[65] Autel Robotics. (2017) X-Star Premium Manual. [Online]. Avail-

able: https://3yn7uf15sq4l1ivh8t2tfir5-wpengine.netdna-ssl.com/wp-content/uploads/
2016/09/x-star-premium-manual.pdf
[66] Blade. (2016) Blade Conspiracy Instruction Manual. [Online]. Available: http:
//www.bladehelis.com/ProdInfo/Files/BLH02000-Manual-EN.pdf
[67] Blade. (2017) Theory XL Instruction Manual. [Online]. Available: http://www.

bladehelis.com/ProdInfo/Files/BLH02150-Manual-EN.pdf
[68] Blade. (2017) Torrent 110 FPV PNP. [Online]. Available: http://www.bladehelis.com/
ProdInfo/Files/BLH04050-Manual-EN.pdf
[69] DJI. (2018) Mavic 2 User Manual. [Online]. Available: https://dl.djicdn.com/

downloads/Mavic_2/20180920/Mavic+2+Pro+Zoom+User+Manual_v1.2.pdf
[70] DJI. (2017) Phantom 4 User Manual. [Online]. Avail-

able: https://dl.djicdn.com/downloads/phantom_4_pro/20171017/Phantom_4_Pro_Pro_
Plus_User_Manual_EN.pdf
[71] DJI. (2018) Matrice 2000 User Manual. [Online]. Available: https://dl.djicdn.com/
downloads/M200/20180910/Matrice_210_210_RTK_User_Manual_EN.pdf
[72] EHang. (2018) GhostDrone page. [Online]. Available: http://www.ehang.com/param.

html
46
[73] EHang. (2018) Falcon page. [Online]. Available: http://www.ehang.com/bat/
[74] FlyAbility. (2018) ELIOS Inspect and Explore. [Online]. Available: https:
//drive.google.com/file/d/1IVuUpzUAz3RtH1HCTxbI4dAgq9H2TCyQ/view
[75] GoPro. Karma Let’s Get Started. [Online]. Avail-

able: https://d3rd8mpr9gp2n3.cloudfront.net/content/dam/post-purchase/karma/pdfs/
Karma_MainSKUQSG_HERO5_Default_LR_REVA.pdf
[76] GoPro. (2018) Karma Tech Specs. [Online]. Available: https://images-eu.

ssl-images-amazon.com/images/I/817tGGpn0KS.pdf
[77] Ubsan. X4 Air User Manual. [Online]. Available: https://www.hubsan.com/support/

files/products/H501A/um/manual-en.pdf
[78] Ubsan. X4 Plus User Manual. [Online]. Available: https://www.hubsan.com/support/

files/products/H107P/um/manual-en.pdf
[79] Ubsan. X4 Pro User Manual. [Online]. Available: https://www.hubsan.com/support/

files/products/H109SS/um/manual-en.pdf
[80] Ideafly. (2015) Storm page. [Online]. Available: http://www.idea-fly.com/En/

parameter.asp?id=6
[81] Ideafly. (2015) Grasshopper page. [Online]. Available: http://www.idea-fly.com/En/

productshow.asp?id=23
[82] Ideafly. (2015) Poseidon page. [Online]. Available: http://www.idea-fly.com/En/

productshow.asp?id=21
[83] Parrot. (2018) Anafi page. [Online]. Available: https://www.parrot.com/us/drones/

anafi#technicals
[84] Parrot. (2018) Bebop User Guide. [Online]. Available: https://www.parrot.com/files/

s3fs-public/firmware/bebop-2_user-guide_uk_2_0.pdf
[85] Parrot. (2018) Disco page. [Online]. Available: https://www.parrot.com/us/drones/

parrot-disco-fpv#technicals
[86] PowerVision. (2018) PowerEgg page. [Online]. Available: https://www.powervision.

me/uk/product/poweregg
[87] PowerVision. (2018) PowerEye page. [Online]. Available: https://www.powervision.

me/uk/product/powereye/specs
[88] Redcat racing. (2017) Carbon 210 Race Drone Manual. [Online]. Available:
http://www.redcatracing.com/manuals/CARBON210MANUAL.pdf
47
[89] SkyDrones. (2018) Zangão Characteristics. [Online]. Available: https://skydrones.
com.br/wp-content/uploads/2017/07/SkyDrones_Zangao.pdf
[90] (2018) Response e-mail. [Online]. Available: https://drive.google.com/file/d/

140ScdddqtOSy45Kso1P_rWWv-YxfdsB-/view
[91] SkyDrones. (2018) Pelicano Characteristics. [Online]. Available: https://skydrones.

com.br/wp-content/uploads/2017/08/SkyDrones_Pelicano.pdf
[92] SkyDrones. (2018) Strix AG Characteristics. [Online]. Available: https://skydrones.

com.br/wp-content/uploads/2017/07/SkyDrones_Strix-AG.pdf
[93] Syma. (2016) X5SW User Manual. [Online]. Available: http://www.symatoys.com/

upload/201609/13/201609131500596655.pdf
[94] Syma. (2018) X8Pro User Manual. [Online]. Available: http://www.symatoys.com/

goodshow/x8-pro-syma-x8-pro-subvert-your-visual-enjoy-the-f.html
[95] Syma. (2018) Z1 User Manual. [Online]. Available: http://www.symatoys.com/

goodshow/z1-syma-z1-enjoy-the-convenience-and-intelligence.html
[96] Traxxas. (2018) Aton page. [Online]. Available: https://traxxas.com/products/landing/

Aton/
[97] Traxxas. (2018) QR-1 site. [Online]. Available: https://traxxas.com/products/models/

heli/6208qr1
[98] Traxxas Latrax. Latrax Alias page. [Online]. Available: https://latrax.com/products/

alias?t=specs
[99] UdiR/C. (2017) Discovery2 U818A page. [Online]. Available: http://www.udirc.com/

products/discovery2-u818a-plus
[100] UdiR/C. (2016) Peregrine U28W Guide. [Online]. Available: http://www.udirc.com/

wp-content/uploads/2016/12/U28W-guide.pdf
[101] UdiR/C. (2017) Predator page. [Online]. Available: http://www.udirc.com/products/

u842wifi
[102] Walkera. (2017) Furious 215 page. [Online]. Available: https://www.walkera.com/

index.php/Goods/canshu/id/45.html
[103] Walkera. (2017) Voyager page. [Online]. Available: http://www.walkera.com/v4/#v42
[104] Walkera. (2017) VITUS Quick Start Guide. [Online]. Available: https://www.
walkera.com/downloads/vitus/VITUS%20Quick%20Start%20Guide-EN-V1.3.pdf
[105] WlToys. Q696-A page. [Online]. Available: https://wltoys.eu/wl_en/catalog/product/

view/id/2353/s/wltoys-q696-a/category/63
48
[106] WlToys. V383 page. [Online]. Available: https://wltoys.eu/wl_en/catalog/product/
view/id/2359/s/wltoys-v383/category/63
[107] WlToys. Q696-D page. [Online]. Available: https://wltoys.eu/wl_en/catalog/product/

view/id/2352/category/63
[108] Yunnec. Typhoon 4K page. [Online]. Available: http://us.yuneec.com/

typhoon-4k-specs
[109] Yunnec. Tornado User Manual. [Online]. Available: http://us.yuneec.com/files/

downloads/h920/h920_user_manual_A5_(st12)_v1.0.pdf
[110] Yunnec. H520 Owners Guide. [Online]. Available: http:

//commercial.yuneec.com/files/downloads/H520/Yuneec%20H520%20Manual+
Datapilot%20Operations%20Guide_EN.pdf
[111] M. Wax and T. Kailath, “Detection of signals by information theoretic criteria,” IEEE
Transactions on Acoustics, Speech, and Signal Processing, vol. 33, no. 2, pp. 387–392,
1985.
[112] J. P. C. L. da Costa, Parameter Estimation Techniques for Multi-Dimensional Array

Signal Processing. Shaker Verlag, 2010.
[113] Z. Chen, G. Gokeda, and Y. Yu, Introduction to Direction-of-Arrival estimation.

Artech House, 2010.
[114] M. A. M. Marinho, F. Antreich, J. P. C. L. da Costa, and J. A. Nossek, “Reduced rank

TLS array interpolation for DOA estimation,” in 18th International ITG Workshop on
Smart Antennas (WSA). VDE, 2014, pp. 1–5.
[115] M. A. Marinho, F. Antreich, J. P. C. L. da Costa, and J. A. Nossek, “A signal adaptive

array interpolation approach with reduced transformation bias for DOA estimation of
highly correlated signals,” in International Conference on Acoustics, Speech and Signal
Processing (ICASSP). IEEE, 2014, pp. 2272–2276.
[116] A. DEVICES. (2017) AD9361 datasheet. [Online]. Available: http://www.analog.

com/media/en/technical-documentation/data-sheets/AD9361.pdf
[117] MATLAB, version 8.5.0.197613 (R2015a). Natick, Massachusetts: The MathWorks

Inc., 2015.
[118] “Data Sheet dual band antenna, 2.4 & 5GHz, PCB.” [Online]. Avail-
able: https://www.te.com/commerce/DocumentDelivery/DDEController?Action=
showdoc&DocId=Data+Sheet%7F2118309-1%7FA%7Fpdf%7FEnglish%7FENG_
DS_2118309-1_A_2118309-1.pdf%7F2118309-1
49
Appendix A
Proposed Extrapolation Algorithm to

find the EFT thresholds for extremely
low Probability of False Alarm
In this Appendix it is proposed an extrapolation algorithm to estimate the thresholds of

the EFT algorithm in cases that the Probability of False Alarm (Pfa ) is extremely low.
Depending on the noise behavior and the parameters of the scenario [REF1, REF2], the
thresholds can be extremely low. Therefore, the computational complexity of (3.16) can be
prohibitive. In order to overcome such limitation, an extrapolation approach to compute the
thresholds for extremely low values of Pfa is proposed.
Since it is desired to estimate values outside the known limits, an extrapolation method
to estimate these values can be used. It is approximated the descending side of the curve as
a decreasing exponential. In order to simplify the approximation, the logarithmic scale as
exemplified in Fig. A.1 is applied .
Given the two known points in Fig. A.1 obtained by Monte Carlo simulations and given
the linear extrapolation in (A.1), it is possible to compute the two unknown constants a and
b.
Pfa = a · ηP + b. (A.1)
The constants a and b are given by (A.2) and (A.3) by using the two known points
(log10 (Pfa1 ), η1 ) and (log10 (Pfa2 ), η2 ).

Pfa2
log10 Pfa1
a= (A.2)
(η2 − η1 )
 
Pfa2
log10 Pfa1
b = log10 (Pfa2 ) − η2 ·   (A.3)
(η2 − η1 )
50
Probability of False Alarm
Linear Extrapolation
Original function
log10(Pfa1)
log10(Pfa2)
η1 η2
Threshold
Figure A.1: Log linear extrapolation based on two points given by Pfa s and Thresholds ηP .
By replacing a and b in (A.1), it is obtained the expression for the η̂P in (A.4).

η2 Pfa2
log10 (Pfa ) − log10 (Pfa2 ) − (η2 −η1 )
· log10 Pfa1
η̂P = · (η2 − η1 ) . (A.4)
log10 PPfa2
fa1
In our Framework in Fig. 4.5, the EFT with a Pfa = 10−263 is set up. For this Pfa , it is
obtained the following values for the thresholds η1 = 1.5810, η2 = 1.7810, and η3 = 2.1840.
Note that there are only three thresholds, since the smallest eigenvalue is assumed as a
noise eigenvalue in the EFT approach. In Figs. A.2(a), A.2(b) and A.2(c), it it depicted the
extrapolation curves for the thresholds η1 , η2 and η3 , respectively.
51
Measured curve
Measured curve
Extrapolated curve
Extrapolated curve
10-50 Extrapolated with correct MO
10-100 10-100
10-150 10-150
Pfa
Pfa
10-2 10-2
-200
10 10-4 10-200 10-4
10-6 10-6
10-250 10-250
10-8 10-8
10-300 10-300
0 0.02 0.04 0 0.02 0.04 0.06
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
η1 η2
(a) (b)
Measured curve
Extrapolated curve
10-100
10-150
Pfa
10-2
-200
10
10-4
-6
10-250 10
10-8
10-300
0 0.02 0.04 0.06 0.08
0 0.5 1 1.5 2 2.5
η3
(c)
Figure A.2: Thresholds for extrapolated data using computation with M=4 and N=41.
52
ATTACHMENT
In the following attachment we describe configurations of the autopilot hardware con-

troller PixHawk [58, 59]. In an easy way, it helps the new researchers to put some VANT to
fly.
53
ATTACHMENT A
PixHawk Controller Setup
To perform these steps you must have installed the Mission Planner [REF5] software and
connect the PixHawk controller to the computer via USB (recommended) or telemetry.
After you open the Mission Planner software and the controller is already connected to
the computer, you must select the port on the computer that is used for communication in the
upper right corner. The figure A.1 present this step.
Figure A.1: Selection of the port on Mission Planner software.
Now we can use Wizard to make the initial settings, and it will guide us through the
process. So, click on "Initial Setup" and then in "Wizard". The first question is about which
frame layout are you using. Select your frame and click on "Next".
Note: If the VANT is connected with USB cable to the PC, take care that the cable does
not swing very much and may disconnect during the processes below. Otherwise the whole
process should be redone from the beginning.
54
Figure A.2: Selection of the frame layout on Mission Planner software.
A.1 Accelerometer Calibration
One of the most important step is the calibration of the accelerometer. This step allows
Pixhawk to know the correct sides and how it is placed in the frame. The Figure A.3 shows
this step. After placing the VANT on a selected position click on "Start" and do what the
software asks for by looking out for the correct positions. Then, click in next when the
process ends.
Figure A.3: Accelerometer Calibration on Mission Planner software.
55
A.2 Compass Calibration
Now it is necessary to calibrate the compass, this is done on all sides similarly as shown
in Section A.1, except that now we need to rotate the UAV 360o on each measured axis. To
guide us in this process, the white dots of the Figure A.4 must be untouched by the red dot
while the UAV is rotated around the axes. When the PixHawk acquire enough samples the
process finishes automatically.
Figure A.4: Compass Calibration on Mission Planner software.
A.3 Battery Monitor
PixHawk needs to measure the amount of charge present and delivered by the battery
throughout the flight experience. This allows better control and avoid accidents due to lack
of battery. Therefore, select the best option according to the hardware present in your UAV
and enter the correct amount of battery capacity.
A.4 Sonar
With the sonar, the controller can control the approach of obstacles. This is an important
item to prevent collisions or uncontrolled landings from occurring. Then, check the box stat-
ing that you have a sonar installed and select which model is present in your UAV, otherwise
leave the box unchecked and click on next.
56
Figure A.5: Battery selection on Mission Planner software.
Figure A.6: Sonar selection on Mission Planner software.
A.5 Radio Calibration
On remotely piloted flights, the commands are sent to the controller via a Radio Control
(RC). This RC must to be calibrated in order to allow the controller to know what are the
values of the limits of the controller. This step is very simple, and what you need to do
is move all the sticks and switches to their limits so the controller can record these values.
Therefore, make sure the radio is turned on and remove the propellers when you are done
with this setting. Click on "Calibrate Radio" to start and, when finished, click on "Click
when done". The boundaries are marked by the red bars in Figure A.7.
57
Figure A.7: Radio calibration on Mission Planner software.
A.6 Fly modes
This chapter is one of the most important in this work. In addition to providing an
overview of the configuration of the controller, this work aims to explain the different flight
modes for when to make your first flights. Knowing which flight mode is most recom-
mended, whether a beginning, intermediate or advanced pilot.
58
The Table A.1 presents the flight modes and its descriptions. This table was built with the modified informations taken from Escola de Drones
[REF6] webpage.
Table A.1: Table of flight modes present on PixHawk.
Flight mode Description

Stabilize The stabilize mode does not have automatic altitude control or positioning (GPS), it is one of the most difficult modes,
since you have to regulate the accelerator constantly, it is a very active piloting and not recommended for beginners,
however, after learning to fly in modes with altitude and positioning controlled automatically it is highly recommended
that you practice Stabilize mode, because in situations that the autopilot fails you must know how to safely drive your
drone to the ground in manual mode.
Alt Hold In Alt Hold mode the acceleration is controlled automatically to maintain the drone at the same altitude, it is an easy
mode of piloting, considering that you will worry more about the right control, leaving the left stick in the middle the
drone will hover at the current altitude detected by the inertial sensors, when you increase the acceleration the drone
rises, then you return the stick to the middle and it maintains the new altitude. In this mode when you land you need to
wait a few seconds to disarm, this is normal, a necessary check to prevent disarmament in the air.
Loiter The Loiter mode is the GPS mode, the VANT will try to maintain its current position, a good GPS signal is essential
for a quiet flight, if there are GPS system failures and the drone starts to move away one must quickly change to the Alt
Hold or Stabilize (depending on your ability) to regain control, some people with less experience turn on failsafe, it is
also an option, but if the home position was recorded with error due to lack of satellites or some problem and the drone
start pulling away, you should get back in control immediately by turning off the fail safe.
Return To Launch (RTL) The RTL mode is operated in conjunction with GPS as well (and depends on a good signal) and has the function of
bringing the drone to the recorded home point, ie the takeoff position. When using this flight mode it is very important
to consider the parameter "RTL_ALT", which is the height the drone will set for the return mission home, by default
the parameter is set to 15M if you want the drone to return at the same height that it is currently this parameter must be
set to 0.
59
Drift In this mode the Roll is automatically controlled and the other controls are available to the pilot, one-handed control
is possible in most of the time, the YAW is not used and you will only use the left control when you want to change
altitude, it is similar to the control of a joystick.
Guided The guided mode is like dynamic missions, instead of using a control, you command the flight via waypoints directly
from the ground station (computer / tablet).
Circle In this mode the drone will be flying in circles, the radius must be defined in the "CIRCLE_RADIUS" parameter. If the
parameter is set to 0 the drone will rotate slowly on the axis itself, good for panning.
Land Land mode is rarely used, it is used to land. For beginners it is recommended if you lose control. But in general, it does
not pay to spend a key position of the radio with him.
Follow Me The Follow Me mode makes the drone follow you, for this you must use a ground station that has GPS (can be an USB
GPS or Bluetooth).
Simple and Super Simple The Simple and Super Simple modes are combined with other flight modes, which serve to make the control always
working from the pilot’s point of view. The simple uses the compass and the super uses the GPS, having as reference
the Home Point.
Auto The auto mode is used when you have previously programmed a waypoint and reached the point at which you want to
execute it. It is essential a good functioning GPS and inertial sensors to perform a mission, it is recommended a small
flight in Loiter mode to test the drone before starting the auto mode, but you can also start the mission on the ground
(accelerator at 0, configured key in "auto", go gradually accelerating, when you get in the middle the drone will start
the mission).
Acro Acro Mode, as the name suggests, is intended for stunts and should only be used by experienced pilots.
60
Now, with the concepts of the Table A.1, you can select six best flight modes, described
in Figure A.8 which best describe your experience. When finished click "Save modes", for
the controller to save the chosen settings.
Figure A.8: Flight modes configuration on Mission Planner software.
A.7 Failsafe
This step configures what happens when one of the three following fails is triggered.
• Battery failsafe: When the battery is finished and the VANT is still on flight.
• Ground Station failsafe: When the VANT lost the telemetry communication.
• Throttle failsafe: When the VANT lost communication with the RC.
Select the options that best describe your needs and then click on "Next"
A.8 GeoFence Settings
This settings configures the radius, altitude and the action for Geofence, that is a virtual
fence, based on home position, that the PixHawk will not fly outside of.
A.9 Possible Errors
Now the settings of the VANT are fully completed. It is ready to fly, however it is
important to know possible problems presented by the Mission Planner [REF5] at the take
61
Figure A.9: Failsafe configuration on Mission Planner software.
Figure A.10: GeoFence configuration on Mission Planner software.
off time. The Table A.2 presents the most common errors and was built with the modified
informations taken from Ardupilot [REF7] webpage.
62
Table A.2: Table of most common errors displayed on Mission Planner software.
Errors Description
RC not calibrated The radio calibration has not been performed. RC3_MIN and
RC3_MAX must have been changed from their default values
(1100 and 1900), and for channels 1 to 4, MIN value must be
1300 or less, and MAX value 1700 or more.
Compass not calibrated The compass(es) has not been calibrated. the COM-
PASS_OFS_X, Y, Z parameters are zero or the number or type
of compasses connected has been changed since the last compass
calibration was performed.
Need 3D Fix The GPS does not have a 3D fix (four satellites) and the vehicle
is in a flight mode that requires the GPS and/or the circular fence
is enabled.
Bad Velocity The vehicle’s velocity (according to inertial navigation system) is
above 50cm/s. Issues that could lead to this include the vehicle
actually moving or being dropped, bad accelerometer calibration,
GPS updating at below the expected 5hz.
High GPS HDOP The GPS’s HDOP value (a measure of the position accuracy) is
above 2.0 and the vehicle is in a flight mode that requires GPS
and/or the circular fence is enabled. This may be resolved by
simply waiting a few minutes, moving to a location with a bet-
ter view of the sky or checking sources of GPS interference (i.e.
FPV equipment) are moved further from the GPS. Alternatively
the check can be relaxed by increasing the GPS_HDOP_GOOD
parameter to 2.2 or 2.5. Worst case, the pilot may disable the
fence and take-off in a mode that does not require the GPS (i.e.
Stabilize, AltHold) and switch into Loiter after arming but this is
not recommended.
Alt disparity The barometer altitude disagrees with the inertial navigation (i.e.
Baro + Accelerometer) altitude estimated by more than 2 me-
ters. This message is normally short-lived and can occur when
the flight controller is first plugged in or if it receives a hard jolt
(i.e. dropped suddenly). If it does not clear, the accelerometers
may need to be calibrated or there may be a barometer hardware
issue.
63
References
[REF1] J. P. C. L. da Costa, Parameter Estimation Techniques for Multi-Dimensional

Array Signal Processing. Shaker Verlag, 2010.
[REF2] J. P. C. L. da Costa, F. Roemer, M. Haardt, and R. T. de Sousa, “Multi-

dimensional model order selection,” EURASIP Journal on Advances in Signal
Processing, vol. 2011, no. 1, p. 26, 2011.
[REF3] L. Meier, P. Tanskanen, L. Heng, G. H. Lee, F. Fraundorfer, and M. Pollefeys,

“Pixhawk: A micro aerial vehicle design for autonomous flight using onboard
computer vision,” Autonomous Robots, vol. 33, no. 1-2, pp. 21–39, 2012.
[REF4] L. Meier, D. Honegger, and M. Pollefeys, “Px4: A node-based multithreaded

open source robotics framework for deeply embedded platforms,” in Robotics
and Automation (ICRA), 2015 IEEE International Conference on. IEEE, 2015,
pp. 6235–6240.
[REF5] Ardupilot. (2016) Software Mission Planner. [Online]. Available: http:

//ardupilot.org/planner/
[REF6] Escola de Drones. (2015) MODOS DE VOO APM E PIXHAWK. [Online].

Available: http://www.escoladedrones.com.br/modos-de-voo-apm-e-pixhawk/
[REF7] Ardupilot. (2016) Failure messages. [Online]. Available: http://ardupilot.org/

copter/docs/prearm_safety_check.html#failure-messages
64

Low Cost Antenna Array Based Drone Tracking Device For Outdoor Environment

Enviado por

Direitos autorais:

Formatos disponíveis

Low Cost Antenna Array Based Drone Tracking Device For Outdoor Environment

Enviado por

Dados do documento

Descrição original:

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Low Cost Antenna Array Based Drone Tracking Device For Outdoor Environment

Enviado por

Direitos autorais:

Formatos disponíveis

LOW COST ANTENNA ARRAY BASED

DRONE TRACKING DEVICE FOR

MARCOS TEIXEIRA DE OLIVEIRA

DISSERTAÇÃO DE MESTRADO EM SISTEMAS MECATRÔNICOS