898 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 22, NO.

2, FEBRUARY 2021

Vehicle Position Correction: A Vehicular Blockchain

Networks-Based GPS Error Sharing Framework
Changle Li , Senior Member, IEEE, Yuchuan Fu , Fei Richard Yu, Fellow, IEEE,
Tom H. Luan , Senior Member, IEEE, and Yao Zhang

Abstract— The positioning accuracy of the existing vehicular and vehicle state data. Vehicles with accurate position infor-
Global Positioning System (GPS) is far from sufficient to support mation can adjust driving conditions according to dynamic
autonomous driving and ITS applications. To remedy that, changes in the driving environment, thus enabling safe driving
leading methods such as ranging and cooperation have improved
the positioning accuracy to varying degrees, but they are still full and improving the driving experience.
of challenges in practical applications. Especially for cooperative Global Positioning System (GPS) is one of the most widely
positioning, in addition to the performance of methods, cooper- used technologies. Although GPS signals are available most of
ators may provide false data due to attacks or selfishness, which the time, accurate positioning is still challenging for supporting
can seriously affect the positioning accuracy. By fully exploiting autonomous driving and other applications where positioning
the characteristics of blockchain and edge computing, this paper
proposes a vehicular blockchain-based secure and efficient GPS accuracy is critical. For instants, in urban canyons, due to
positioning error evolution sharing framework, which improves multipath interference, GPS accuracy in urban canyons ranges
vehicle positioning accuracy from ensuring security and credibil- between 30-50 meters [1]. Even under normal circumstances,
ity of cooperators and data. First, by analyzing the GPS error, GPS positioning errors may lead to wrong navigation, waste
a bridge can be established between the sensor-rich vehicles of time, and traffic accidents. Fortunately, for some vehicles
and the common vehicles to achieve cooperation by sharing
the positioning error evolution at a specific time and location. with multiple sensors (hereinafter referred to as sensor-rich
Particularly, the positioning error evolution is obtained by a deep vehicles), such as camera and Lidar, other sensing methods
neural network (DNN)-based prediction algorithm running on the can be used to improve their positioning accuracy [2], [3].
edge server. We further propose to use blockchain technology for However, on the one hand, for current traffic scenarios, most
storage and sharing the evolution of positioning errors, mainly common vehicles still exist, and it is impossible to obtain
to guarantee the security of cooperative vehicles and mobile
edge computing nodes (MECNs). In addition, the corresponding accurate sensing information. On the other hand, Lidar sensors
smart contracts are designed to automate and efficiently perform are expensive, which restricts their large-scale application.
storage and sharing tasks as well as solve inconsistencies in time Therefore, how to improve GPS positioning accuracy is still
scales. Extensive simulations based on actual data indicate the worth studying.
accuracy and security of our proposal in terms of positioning A lot of work has been proposed to improve the GPS
error correction and data sharing.
positioning accuracy of vehicles, from hybrid systems,
Index Terms— Positioning error evolution, edge computing, infrastructure-based positioning, vehicle cooperative
DNN, vehicular blockchain, smart contract. positioning (VCP), cooperative positioning (CP) systems,
fusion systems, to mobile phone-based positioning, etc.
I. I NTRODUCTION
Demetriou et al. [1] proposes a system called CoDriv for

P OSITIONING information is very important for vehicles,

especially for autonomous vehicles, which can be used to
navigate in real-time with other data such as geographic data
providing the accurate position of sensor-rich vehicles to
a common vehicle. Angelis et al. [4] propose to use the
Global Navigation Satellite System (GNSS) and cellular
Manuscript received February 6, 2019; revised May 31, 2019 and networks to form a hybrid scheme for user positioning in
October 8, 2019; accepted December 16, 2019. Date of publication Jan- urban scenarios. Alam and Dempster [5] discuss modern
uary 3, 2020; date of current version February 2, 2021. This work was and conventional CP systems. Yassin and Rachid [6]
supported in part by the National Natural Science Foundation of China under
Grant U1801266, Grant 61571350, and Grant 61601344, in part by the Key explain the principles behind positioning techniques used in
Research and Development Program of Shaanxi under Contract 2018ZDXM- satellite networks, mobile networks, and Wireless Local Area
GY-038 and Contract 2018ZDCXL-GY-04-02, in part by the Youth Innovation Networks. Some infrastructure-based positioning methods are
Team of Shaanxi Universities, and in part by the Science and Technology
Projects of Xi’an, China, under Grant 201809170CX11JC12. The Associate proposed (e.g., anchor nodes [7], and magnetic positioning
Editor for this article was S.-H. Kong. (Corresponding author: Changle Li.) system [8]). Wang et al. [9] design a dedicated short-
C. Li, Y. Fu, and Y. Zhang are with the State Key Laboratory of Inte- range communications (DSRC)-based vehicle cooperative
grated Services Networks, Xidian University, Xi’an 710071, China (e-mail:
clli@mail.xidian.edu.cn). positioning enhancement system. Li et al. [10] develop
F. R. Yu is with the Department of Systems and Computer Engineering, roadside equipment (RSE)-assisted lane-level positioning
Carleton University, Ottawa, ON K1S 5B6, Canada. method using GPS data and received signal strength (RSS)
T. H. Luan is with the School of Cyber Engineering, Xidian University,
Xi’an 710071, China. data. Walter et al. [11] propose to use the sensors of Android
Digital Object Identifier 10.1109/TITS.2019.2961400 smartphone for vehicle navigation when the GPS signal is lost.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
LI et al.: VEHICLE POSITION CORRECTION: A VEHICULAR BLOCKCHAIN NETWORKS-BASED GPS ERROR SHARING FRAMEWORK 899

These efforts use different methods to improve vehicle and security of our proposal in terms of positioning error
positioning accuracy to degrees. However, the methods based prediction, error correction and data sharing process.
on expensive sensors are still challenging for applying to The remainder of the paper is organized as follows. The
common vehicles. The ranging based method also has errors related work is presented in Section II. In Section III, we depict
in the ranging process [10]. Therefore, starting from the the system architecture and analyze the GPS error. Based on
GPS positioning error characteristics, it is very meaningful the system model, the deep learning algorithm for analyzing
to study a scheme that can adapt to the current scenario of GPS error evolution is proposed in Section IV. In Section V,
the coexistence of common vehicles and sensor-rich vehicles we present the vehicular blockchain structure and smart con-
to facilitate positioning error correction. On the one hand, tracts. Extensive simulation results are discussed in Section VI,
the mobile edge computing nodes (MECNs) can store accurate and conclusion is drawn in Section VII.
map information, and sensor-rich vehicles can easily obtain
accurate positioning. On the other hand, thanks to the comput- II. R ELATED W ORK
ing power and storage capabilities of edge computing, sensor-
In this section, we survey the existing works on the improve-
rich vehicles participating in cooperative positioning can store
ment of vehicle positioning accuracy and vehicular blockchain.
the positioning error model in MECNs to provide a reference
for common vehicles. However, due to attacks or selfish
reasons, cooperative vehicles and MECNs cannot be fully A. Positioning of Vehicles
trusted, which may provide false data. How to improve the There are two main types of methods for positioning of
security and reliability of cooperators and data is worth pon- vehicles. One is a non-cooperative positioning method based
dering. Indeed, blockchain is a very promising technology for on the vehicle’s own body sensor such as radar, camera,
secure storing and sharing data in many application scenarios, and Lidar, and the other is a CP method [15], [16]. For
such as the Internet of Things (IoT) and vehicular networks non-cooperative positioning, in most cases, the positioning
[12], [13]. Applying blockchain technology for vehicle posi- accuracy is relatively accurate. Chen et al. [17] propose a
tion information sharing can bring many benefits, such as secu- positioning algorithm based on machine vision that utilizes
rity, automation, and transparency [14]. Therefore, positioning low-cost monocular cameras. In [18] and [19], Lidar is adopted
accuracy can be improved by avoiding malicious false data and to observe the surrounding environment and match the obser-
motivating more users to participate in cooperative positioning. vation with a priori knew 3D point cloud map for estimating
This work proposes a vehicular blockchain-based frame- the position of the vehicle within the map. However, this
work for improving GPS positioning accuracy by ensuring sensor-based approach has the drawback of a line-of-sight
the security and credibility of cooperative data. In particular, characteristic and severely affected by the weather. The authors
as information providers, sensor-rich vehicles provide posi- of [20] propose a sensor fusion-based vehicle positioning sys-
tioning errors obtained by other methods to common vehicles tem by using GPS, camera and digital map, etc., which costs
(i.e., data requesters). Specifically, the characteristics of GPS less. High definition (HD) map contains the accurate three-
error are analyzed and the GPS positioning error evolution dimensional characterization of the road network (centimeter-
model is obtained by the deep neural networks (DNN). Then level accuracy). However, the requirements of high accuracy
we put forward a GPS error evolution sharing framework that and real-time also make the collecting and processing of large
utilizes edge computing and blockchain technology. In order amounts of data become extremely challenging. At the same
to solve the problem of inconsistent block generation time time, its huge amount of data consumption of communication
and information exchange time scale, we designed smart and computing resources should not be underestimated.
contracts that run the data storage and sharing process securely For CP, it is through direct communication between several
and efficiently. The main contributions of this paper are vehicles to improve their positioning accuracy [5]. The authors
four-fold: of [5] discuss modern and conventional CP systems in detail.
• System Solution: A vehicular blockchain-based GPS error One is based on the method of ranging, such as RSS, time of
evolution model sharing framework is proposed. The arrival (TOA), angle of arrival (AOA), and Time Difference-of-
corresponding smart contracts for data storage and shar- Arrival (TDOA), which estimate the relative position directly
ing are designed. Our proposal improves the accuracy at the RF-signal level [16]. However, these measurements are
of cooperative positioning by ensuring the security and challenged by some drawbacks varying from complexities of
reliability of cooperators and data. the time-synchronization, occupations of the high-bandwidth,
• AI-based Algorithm: We propose a DNN-based error to huge costs on the implementations. Kloeden et al. design
correction algorithm that runs on the edge server and can a low-cost prototype system for vehicle self-localization using
take advantage of the computing power and low latency AOA technique [21], which can achieve lane positioning.
characteristics of edge computing. In order to solve the problem of inaccurate positioning caused
• Collective Learning Strategy: Cooperation between vehi- by nonline-of-sight (NLoS), a TOA-based positioning technol-
cles by sharing DNN models instead of positioning error ogy is proposed [22]. In addition to the ranging-based method,
data, which makes the results more adaptive to the current the other is to exchange positioning data, speed and other
driving environment. information between vehicles to improve positioning accuracy.
• Validation: The developed proposal is evaluated by exten- Cooperative vehicle communication technology offers new
sive simulations. Simulation results validate the accuracy opportunities for CP methods [9], [23]–[25]. The authors
900 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 22, NO. 2, FEBRUARY 2021

of [9] take use of DSRC for sharing motion states and which is often used between organizations. In the private
physical measurements among vehicles to improve vehicle blockchain, all nodes in the network are in the hands of an
positioning accuracy. In [23] and [24], GPS position estimates organization or entity.
and vehicle-to-vehicle (V2V) measurements are combined for In the blockchain system, non-tamperable data structures,
positioning enhancement. The authors also explain the effect distributed network consensus mechanisms, and increasingly
of noise on the GPS and received signal strength of V2V. flexible smart contracts are representative innovations. The
Kakkavas et al. [25] investigate the performance of relative core problem of distributed networks is how to achieve consen-
posting using V2V and the importance of AOA and TDOA for sus efficiently [14]. The consensus mechanisms of the current
location estimation. A drawback of this type of method is that blockchain system are mainly the following: Practical Byzan-
the exchange of information between vehicles is unreliable, tine Fault Tolerance (PBFT), Proof of Work (PoW), Proof of
which results in reduced awareness of the relative position of Stake (PoS), and Delegated Proof of Stake (DPoS) [33]–[35].
the own vehicle relative to the surrounding vehicle. Cooper- The smart contract of blockchain technology is the decen-
ators who are not fully trusted will also affect the accuracy tralized, information-sharing program code deployed on the
of cooperative positioning. These inspire our work to improve blockchain. The parties who signed the contract agreed on
positioning accuracy from the perspective of cooperators and the contract content and deployed it on the blockchain in the
cooperative data security. form of smart contracts, which can automatically execute the
There are also some methods based on smartphones, road- contract on behalf of the signatories without relying on any
side unit (RSU), and other infrastructures. The former dra- central organization [14], [36]. The operation of the smart con-
matically depends on the accuracy in real-time event data tract is as follows. The smart contract encapsulates a number
provided by smartphones. Both [11] and [26] use smartphones of predefined states, conversion rules, trigger conditions, and
to improve GPS positioning accuracy. In [11], the sensor corresponding operations. After being signed by the parties,
data of the smartphone is used to fuse the GPS data of the it is attached to the blockchain data in the form of program
vehicle to realize the vehicle navigation when the GPS is not code, and is recorded after propagation and verification by the
available. In order to solve the problem of insufficient accu- blockchain network. In the distributed ledger of each node,
racy of the current positioning technology, the authors [26] the blockchain can monitor the status of the entire smart
introduce a smartphone-based lane detection system, which contract in real-time and activate and execute the contract
uses the low-power inertial sensor data in the smartphone after confirming that certain trigger conditions are satisfied.
to provide an accurate estimate of the current lane position Smart contracts encapsulate the complex behavior of nodes
of vehicles. In [10], the RSE is fully utilized for lane-level in a blockchain network, providing a convenient interface
positioning. Both GPS and RSS data are considered and for building upper-layer applications based on blockchain
high positioning accuracy is achieved by connected vehicle technology.
networks. A hybrid TDOA and AOA localization method
in three-dimensional (3D) space are proposed in [27]. The
principle of infrastructure-assistant methods are the same as C. Blockchain for Internet of Vehicles
the ranging method of CP, but the positioning is accurate At present, the blockchain technology has not been applied
when the vehicle needs to access multiple infrastructures to vehicle positioning, so we investigate the state-of-the-art
at the same time. The focus of both [28] and [29] is on work in applying blockchain to the Internet of Vehicles.
machine learning algorithms for the indoor positioning system. Kang et al. [12] make use of blockchain technology to
By contrast, the focus of this paper is on the use of vehicular achieve secure data sharing in the Internet of Vehicles. In this
blockchain technology to improve the security and reliability paper, the vehicle can be a data provider or requester, using
of cooperators and cooperative data for enhancing the position- efficient contracts running on the blockchain for efficient data
ing accuracy in outdoor vehicular environments. These work storage and sharing. In [37], a distributed and self-managed
[28], [29] demonstrate the feasibility of applying machine Vehicle Ad-hoc Networks (VANETs) architecture based on
learning algorithms for positioning. concept of Ethereum is proposed. Specifically, Ethereum’s
contract system is utilized to organize and self-manage various
applications running on the RSU to provide different services
B. Background of Blockchain Technology to the vehicle. In order to protect the vehicle from security
Blockchain is a decentralized infrastructure widely used and privacy threats, a blockchain-based system is proposed
in emerging digital cryptocurrencies. It has the characteris- to increase the security of the vehicular ecosystem [38].
tics of decentralization, data not tampering, and trustwor- Singh and Kim [39] use the blockchain technology applied
thiness [14]. According to the degree of centralization of in Bit-coin to propose an Intelligent Vehicle (IV) data
the blockchain systems, it is mainly divided into three cate- sharing framework to build trust and reliability. In the trust
gories: public blockchain, consortium blockchain, and private environment based IV framework, the blockchain is used
blockchain [30]. The public blockchain allows any node to as the backbone of the IV data sharing. Yang et al. [40]
join the blockchain network and view information on the propose a decentralized trust management system in vehicular
blockchain, such as Bitcoin [31] and Ethereum [32]. The networks based on blockchain technology to solve the
consortium blockchain allows authorized nodes to join the problem that the vehicle is difficult to evaluate the credibility
network and can view information according to permissions, of the received message due to the untrusted environment.
LI et al.: VEHICLE POSITION CORRECTION: A VEHICULAR BLOCKCHAIN NETWORKS-BASED GPS ERROR SHARING FRAMEWORK 901

TABLE I TABLE II
M AIN N OTATIONS M AIN A BBREVIATIONS

In this system, the vehicle can use a Bayesian inference

model to verify received messages from neighboring vehicles.
However, although these important works have made great
efforts to improve the accuracy of vehicle positioning, there
are still challenges in practical applications. In detail, in the
current mixed traffic scene, GPS is still the main way of
positioning most vehicles. It is meaningful to improve the
positioning accuracy by analyzing the GPS positioning error
characteristics to carry out corresponding research. In addition
to the positioning method itself, data security is an important
factor to consider in cooperative positioning methods because
the cooperative vehicle and MECNs are not completely reli-
able. Therefore, this paper uses DNN and vehicular blockchain
technology to improve the positioning accuracy of vehicles to
take a further step in terms of accuracy and security.

III. S YSTEM M ODEL

In this section, we describe the system architecture and the
models that need to be used for positioning error correction.
In addition, we analyze the possibility of sharing vehicle
Fig. 1. System architecture of the GPS error sharing framework using
positioning error models for improving positioning accuracy. vehicular blockchain.
The main notices and abbreviations in this article are listed
in Table I and Table II.
When a sensor-rich vehicle approaches the landmark,
A. Positioning Scenarios and System Architecture the distance and angle between itself and the landmark can
be obtained by on-board sensors. And the position of the
Generally speaking, there are two types of vehicles, i.e., landmark can be used to calculate the current position of
common vehicles and sensor-rich vehicles, driving on the the vehicle. For vehicle i , the GPS position is expressed as
road. Among them, common vehicles can only obtain position Pi =< pie , pin >, and the vehicle position Pi =< pie , pin >
information through GPS. In addition to GPS, sensor-rich measured with landmark l can be expressed as:
vehicles can also be assisted by other on-board sensors, such as 
camera and Lidar. Assume that the GPS receivers of the two pie = ple + dil ∗ cos θil
, (1)
types of vehicles are the same. All vehicles can access the pin = pln + dil ∗ sin θil
MECNs to request service within the communication range.
The MECNs have enough storage space and computing power, where Pl =< ple , pln > is the position of landmark l,
which store accurate location information of the infrastructure dil and θil are the distance and angle between vehicle i and
in the area (e.g., landmark), as shown in Fig. 1. landmark l, respectively.
902 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 22, NO. 2, FEBRUARY 2021

Currently, on-board sensors have relative high range accu- to be almost the same, that is, E si ≈ E s j . Bring it into (5) to
racy in most cases. Therefore, for vehicle i , we consider that get:
Pi is the accurate position information compared to Pi . The
 E i j ≤ Eri − Er j  . (6)
GPS positioning error E i can be calculated as:
Since random errors Er are quite small compared to sys-
E i = Pi − Pi . (2) tematic errors E s , E i ≈ E j can be considered in this case.

B. GPS Error Analysis IV. P OSITIONING E RROR C ORRECTION A LGORITHM

In practice, GPS measurements contain a variety of errors, As outlined above, sensor-rich vehicles can utilize other
which can be classified into three categories based on the sensors to correct their position through the location of the
source of the error: 1) errors associated with GPS satellites; landmarks (e.g., stop sign) and get the current GPS position-
2) errors associated with signal propagation; 3) and receiver- ing error. However, due to the discontinuity of landmarks,
related errors. Many studies and our previous work have shown accurate position information based on sensor ranging can-
that according to the nature of the error, these errors can be not be obtained without the landmark assistance. Therefore,
divided into two types: systematic error and random error the positioning error of these road sections can be compensated
[41], [42]. The GPS errors E can be decomposed into: through prediction. In this section, we propose a deep learning-
based algorithm for predicting error evolution that can be used
E = E s + Er , (3) to fill the error of these discontinuous road sections.

where E s and Er are the systematic errors and random error,

A. DNN for Positioning Error Prediction
respectively. Particularly, the systematic error mainly includes
satellite orbit error, satellite clock error, ionospheric delay, DNN is highly fault-tolerant and self-learning, self-
tropospheric delay, receiver clock error and receiver position organizing and adaptive, capable of simulating complex non-
error, providing the same error magnitude and direction for linear mapping. Many factors are affecting the accuracy of
each GPS receiver within a certain range, about 50 − 200km. GPS positioning and the relationships are complicated. There-
The random error mainly includes multipath effect error and fore, the characteristics of the DNN and powerful nonlinear
receiver noise. Usually, random error Er is much smaller than processing capabilities are well suited for predictive problems
the systematic error E s [43], [44]. in traffic scenarios. Due to the powerful computing capacity
For vehicles i and j that are driving on the same road at of the edge server and improved data set, the DNN algorithm
similar times, their relative error  E i j  is satisfied: can run on the edge server for prediction when the landmark
is unavailable.
 E i j  =  (Pi − P j )  −  (Pi − P j )  In our work, a DNN algorithm is proposed to learn and
≤  (Pi − P j ) − (Pi − P j )  compensate for the positioning error. When there is a landmark
on the road, the positioning error obtained by the sensor-rich
≤  E i − E j , (4)
vehicles can be used as the training data of the neural network.
where Pi , P j , Pi , P j , E i and E j are the GPS position, accurate When there is no landmark on the road, the positioning error
position and GPS positioning error of vehicle i and vehicle j , can be predicted by using the GPS data of the vehicle and the
respectively. According to (3) and (4), we can get: vehicle motion state data.
1) Input Nodes: In this scenario, there are five input nodes,
 E i j  ≤  (E si + Eri ) − (E s j + Er j )  i.e., vehicle’s speed v i , vehicle’s acceleration ai , driving direc-
≤  E si − E s j  +  Eri − Er j , (5) tion dri , the GPS position pie and pin . Therefore, the input
vector can be expressed as [v i ai dri pie pin ].
where E si , E s j , Eri and Er j are systematic errors and random 2) Output Nodes: There are two output nodes, which are the
error of GPS positioning of vehicle i and vehicle j , respec- positioning error in the east direction E ie and the positioning
tively. error in the north direction E in . The input vector can be
For the same type of GPS receiver, the positioning error expressed as [E ie E in ].
radius is fixed. In addition, at similar moments, the satel- 3) Layers of Network: In general, a DNN is a neural network
lite combinations that can be observed by each vehicle are with multiple hidden layers between input and output [45].
basically the same. Therefore, when using the same satellite When using the DNN for prediction, the more hidden layer,
combination for positioning, the satellite position deviation the higher the accuracy of the network, while the complexity
caused by satellite clock difference, atmospheric delay, etc., and time-consuming will increase. There is currently no defin-
is almost the same. Moreover, for two vehicles with similar itive formula to calculate the number of network layers, and it
positions, the difference in position between vehicles is negli- is necessary to continuously learn and train according to the
gible relative to the distance from the vehicle to the satellite. data set and specific problems.
In this paper, vehicles for cooperative positioning and common 4) Hidden Layers Nodes: On the one hand, too few hidden
vehicles are considered to be on the same road segment, that layer nodes will make the network unable to have the nec-
is, between two intersections, and generally less than 2 km. essary learning ability and information processing capability.
Therefore, in this case, the systematic error can be considered Conversely, too many hidden layer nodes will not only greatly
LI et al.: VEHICLE POSITION CORRECTION: A VEHICULAR BLOCKCHAIN NETWORKS-BASED GPS ERROR SHARING FRAMEWORK 903

where x and y are the input and output vectors, respectively.

θ is the parameter vector of the loss function. W and b are the
linear relationship coefficient matrix and bias vector of each
hidden layer and output layer. n is the number of samples, ŷ
and yi represent the actual output and desired output of i ’th
sample, respectively.
Therefore, for each sample, f = 21  ŷ − y22 . Suppose the
network has L layers, we can get ŷ = a L . Combine (9) to
solve the gradient of the output layer (i.e., L’th layer) based
on the loss function:
Fig. 2. The deep neural network for predicting error evolution. ∂ f (θ, x, y) ∂ f ∂z L
= L = (a L − y)(a L−1)T  σ  (z L ), (12)
∂W L ∂z ∂ W L
increase the complexity of the network structure, but also and
reduce the speed of network learning, and the network is ∂ f (θ, x, y) ∂ f ∂z L
more likely to fall into local optimum in the learning process. = = (a L − y)  σ  (z L ), (13)
∂b L ∂z L ∂b L
In general, the number of hidden layer nodes are related to the
number of training set samples, the number of input nodes, and where  means Hadamard product.
the number of output nodes. The upper bound of hidden layers For hidden layers, we can obtain
nodes that do not cause overfitting is [46]:
∂ f (θ, x, y) ∂ f (θ, x, y) ∂z L ∂z l+1
S = ··· . (14)
Nh = , (7) ∂z l ∂z L ∂z L−1 ∂z l
(λ ∗ (I + O))
where I and Oare the number of input and output nodes, Combining (8) and (14), the W l and bl of the l’th layer are
respectively, S means number of samples in the training data easy to derive:
set, and λ is a scale factor, typically a constant within [2, 10].
∂ f (θ, x, y) ∂ f ∂z l ∂f
According to Andrew’s course [47], when the number of = = l (a l−1 )T , (15)
hidden layers is greater than 1, the number of nodes of each ∂W l ∂z ∂ W
l l ∂b
hidden layer should be the same. A typical diagram of DNN and
used in this work is shown in Fig. 2.
∂ f (θ, x, y) ∂ f ∂z l
Let z l , a l and bl represent the vector of all inputs, the vector = . (16)
∂bl ∂z l ∂bl
of all outputs, and the vector of all offsets of the l’th layer,
respectively. Thus, the relationship between z l and a l−1 can According to (8) and (9), the relationship between z l and
be expressed as: z l+1
can be represented as
z l = W l ∗ a l−1 + bl , (8)
z l+1 = W l+1 ∗ a l + bl+1
where W l represents the weight from a l−1 connected to z l . = W l+1 ∗ σ (z l ) + bl+1 . (17)
We can also describe their relationship as a vector form:
Therefore, we can obtain
a l = σ (z l )
= σ (W l ∗ a l−1 + bl ), ∂ f (θ, x, y) ∂f
(9) = (W l+1 )T l+1  σ  (z l ), (18)
∂b l ∂b
where σ (·) is activation function.
and
∂ f (θ, x, y) ∂f
B. DNN Training and Parameter Learning = (W l+1 )T l+1  σ  (z l )(a l−1 )T . (19)
∂W l ∂b
The activation functions used by the hidden layers and
output layer of the network are the ReLU function σ (z) = Parameter adjustment according to the following equations
max(z, 0) and Sigmoid function σ (z) = 1+e1 −z , respectively. during the learning process:
The loss function used in this article is mean square error
∂ f l−1 T
(MSE), which is suitable for numerical prediction problems Wl = Wl − η ∗ (a ) (20)
and can be calculated as follows: ∂bl
and
1
n
f (θ, x, y) = ( ŷ − yi )2 , (10) ∂f
n bl = bl − η ∗ , (21)
i=1
∂bl
and
where η is learning rate. The detailed prediction process of
θ = {W, b}, (11) DNN is shown in Algorithm 1.
904 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 22, NO. 2, FEBRUARY 2021

Algorithm 1 Algorithm of Training DNN penalized or removed from the vehicular blockchain. This will
Input: The training samples {(x 1 , y1 ), (x 2 , y2 ), . . . , (x n , yn )}, incent more vehicles to provide information and ensure the
number of layers L, maximum number of iterations Imax , accuracy of the data.
iteration step size s, and stop iteration threshold ε;
Output: The linear relationship coefficient matrix W and bias V. B LOCKCHAIN C ONSTRUCTION
vector b of each hidden layer and output layer; In this section, we introduce the blockchain-based posi-
1: Initialize the values of W and b as random values; tioning error evolution model sharing method, the system
2: for i ter = 1 : Imax do architecture as well as the running smart contracts.
3: for i = 1 : n do
4: Calculate the forward process of each data and retain
the value of the activation function for each layer ail ; A. Vehicular Blockchain
5: Calculate the loss value by loss function f (θ, x, y); The sensor-rich vehicle can get its own positioning error
∂f
6: Calculate the gradient of the output layer ∂z L = evolution about the current road segment. Through the previ-
i
∂f ous analysis, this information can be shared with other vehicles
 σ  (z iL );
∂aiL in similar locations in the near future. In this paper, error
7: for l = 2 : L do correction can be done in two ways: One is to obtain the
∂f l+1 T ∂ f
8: Calculate ∂z l = (Wi ) l+1  σ  (z li ); positioning error evolution of sensor-rich vehicles and directly
∂z i
9: end for combine the GPS positioning information of the vehicle to
10: for l = 2 : L do obtain the corrected position. The other is to predict the
11: Update W l andbl by positioning error of a common vehicle through the error
12: W l = W l − η ni=1 ∂ fl (ail−1 )T ; prediction model obtained by sensor-rich vehicles, thereby
∂z i

13: bl = bl − η ni=1 ∂ fl ; performing error correction. Considering the timeliness of
∂z i information and the validity of location, MECNs can serve
14: end for
as a storage and sharing center for information. However,
15: if Changes in W and b less than ε then
vehicles may provide incorrect data for various reasons (even
16: Stop training.
selfishness), and MECNs may be attacked rather than fully
17: end if
reliable. Therefore, blockchain technology can be used to
18: end for
achieve secure and reliable storage and sharing of information.
19: end for
Since GPS information and positioning error information
can only be applied to vehicles within the area, we adopt a
consortium blockchain based on vDLT blockchain platform,
C. Positioning Error Correction which is faster than the traditional public blockchain and
In this paper, other vehicles, especially common vehicles, the cost is greatly reduced. The vehicular blockchain system
can use the error evolution of blockchain in sensor-rich vehi- architecture is shown in Fig. 1, which is consisted of vehicle
cles in the near future to correct their positioning errors. This nodes, MECNs. The consensus process is controlled by pre-
subsection proposes a strategy based on transfer learning to selected nodes (MECNs). In this scenario, sensor-rich vehicles
perform positioning error correction. are mainly data providers, and common vehicles are mainly
For sensor-rich vehicles, the DNN-based prediction model data requesters. The data provider only encrypts and transmits
and the resulting error evolution results are stored on the the data and transmits the data as a blockchain transaction to
blockchain. The specific data storage and sharing process the nearby MECN for a corresponding reward. In particular,
based on vehicular blockchain will be introduced in Section V. MECNs are not only used for data storage but also perform
For the common vehicle i , after obtaining the model para- the smart contract as controllers during data sharing.
meters and positioning errors of other J vehicles, it first
judges whether or not to use the information. By com-
B. Smart Contracts for Information Sharing
paring the positioning error of vehicle j, j ∈ J (i.e.,
E j = (E ej , E nj )) and average error other J vehicles Smart contracts based on blockchain technology can not
1 J e 1 J
only take advantage of the cost-efficiency of smart contracts
E ave = ( j =1 E j , j =1 E j ). For each vehicle j ,
n
but also avoid the interference of malicious behaviors on the
J J
if ||E j , E ave || ≥ δ, where δ is a certain threshold, it is normal execution of contracts, so that information storage
considered to provide erroneous data and the data is not used. and sharing based on smart contracts can run efficiently and
After culling the erroneous data, the model parameters and safely. In addition, since the block generation period (0.5s)
positioning errors of other vehicles are averaged separately. is inconsistent with the time scale of the vehicle information
Then, vehicle i adds a hidden layer to the neural network and broadcast cycle (200ms), smart contracts can also be used to
continues training, thereby obtaining error evolution, which solve this problem.
can be used to correct its positioning. The newly concluded contract collection will spread to the
When vehicle i uses data stored in the blockchain by other entire network in the form of blocks. As shown in Fig. 3, each
vehicles, it also needs to pay the corresponding fee. Vehicles block contains the following information: hash of the current
that provide multiple error information will be financially block, hash of the previous block, the timestamp when the
LI et al.: VEHICLE POSITION CORRECTION: A VEHICULAR BLOCKCHAIN NETWORKS-BASED GPS ERROR SHARING FRAMEWORK 905

Fig. 3. Smart contract recorded and data recorded in blocks of a blockchain.

consensus was reached, and transactions records set. A block

mainly contains three types of records, that is, data storage
records, data index records, and data sharing records.
Fig. 4. Smart contract process for data storage.
(1) Smart Contract for Recording (SCR): Before performing
a smart contract for data sharing, the data provider needs to
store the encrypted data to the system. We design a storage
solution based on the InterPlanetary File System (IPFS) [48].
IPFS is an emerging standard for storing content addressable
files. Content-addressed storage is a mechanism for storing
information that can be retrieved based on its content rather
than its location. The SCR is deployed on the blockchain,
in which all files are linked to IPFS files. The file link content
mainly includes the local home directory path, the file name
(i.e. hash of the data), and the file size:
File Li nk = (/ home/|| f ile name|| f ile si ze)
The request message for recording data from vehicle v i to a
pu
nearby MECN M EC Nm , which contains the identifier K i ,
the hash of the latest ined block h(Blockt −1), the file link Fig. 5. The consensus process for SCR.
File Li nki and timestamp, that is:
⎧ pu ⎫
⎪
⎪ Ki ⎪
⎪
⎨ ⎬
h(Blockt −1) executed and synchronize to the blockchain, indicating that the
D_Reqv i →M EC Nm : .
⎪
⎪ File Li nki ⎪
⎪ file has been saved successfully, as shown in Fig. 4.
⎩ ⎭
ti mestamp K v The consensus process for SCR is based on PBFT [34],
i
which has been widely used in many practical scenarios.
Particularly, the signature of requester’s private key K iv is In addition, the consensus nodes are obtained by random
used to guarantee the truthfulness of the requested information. calculation, and at least 3 f +1 MECNs are randomly selected,
Then v i will send the encrypted shared data to M EC Nm . After where f ≥ 1. The MECN calculates whether it is selected
the M EC Nm receives the request message and data, it will according to its private key, and broadcasts the result to other
verify that the hash of file content and file size are correct. MECNs. This method ensures that the nodes selected in each
Once the verification is successful, M EC Nm will wrap the round are random and different, which increases security. The
received request into the new record message and add other consensus process mainly has the following steps as depicted
pu
information such as identifier K j , the file link File Li nki in Fig. 5 and Fig. 6:
and timestamp. Similarly, the signature K jM EC N are meant • Broadcast: v i sends request to M EC Nm (consensus
to ensure truthfulness and allow anyone to verify that the nodes or accounting nodes). For accounting nodes,
information has been tampered with. The record message can it needs to forward the transaction to the consensus nodes,
be expressed as: then the consensus nodes broadcast the transaction to
⎧ ⎫
⎪ D_Reqv i →M EC Nm ⎪ other consensus nodes. M EC Nm sends broadcast the
⎪
⎨ pu ⎪
⎬
Km event in Fig. 5 and Fig. 6.
D_ Recv i →M EC Nm : . • Pre-Prepare: The primary (MECN-leader) broadcasts the
⎪
⎪ File Li nki ⎪
⎪
⎩ ⎭ Pre-Prepare message. The MECN_leader (primary node)
ti mestamp K M EC N m will validate the transaction immediately after packing
In case the file is correctly verified, M EC Nm broadcasts the transactions, and then include the validation result
the data record to other MECNs acting as verification nodes to into the Pre-Prepare message for the whole network
verify the signature to ensure its validity. After the consensus broadcast. Pre-Prepare message contains both the ordered
on the event is achieved, the smart contract will be successfully Transaction information and block validation result.
906 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 22, NO. 2, FEBRUARY 2021

Similar to data recording, v i sends the request for recording

the data index to the nearby MECN M EC Nm , that is
⎧ pu ⎫
⎪
⎪ Ki ⎪
⎪
⎨ ⎬
h(Blockt −1)
D I _Reqv i →M EC Nm : .
⎪
⎪ Data i ndex i ⎪
⎪
⎩ ⎭
ti mestamp K v
i

After verification, M EC Nm will wrap the received request

into a new data index record message
⎧ ⎫
⎪
⎪ D I _ Reqv i →M EC Nm ⎪
⎪
⎨ pu ⎬
Km
D I _Recv i →M EC Nm : .
⎪
⎪ Data i ndex i ⎪
⎪
⎩ ⎭
ti mestamp K M EC N
m

Similarly, during the period in which the next block is

Fig. 6. The consensus mechanisms for SCR. generated, each MECN obtains a data index request from
the sensor-rich vehicles within the communication coverage.
The process of adding data index records to the blockchain is
• Prepare: Replicas send prepare message to others. After similar to data recording.
receiving a Pre-Prepare message from MECN_leader, (3) Smart Contract for Sharing (SCS): The common vehicle
a replica checks the validity of the message. When (data requester) v j gets the latest block from the blockchain
the replica discovers that the message is from the and selects the data that needs to be requested to be shared
MECN_leader and receives it for the first time, the replica according to its own situation and data description. When v j
will broadcast a Prepare message indicating consistency issues a request for location information sharing, the SCS
with the result of the MECN_leader. is triggered. Assuming v j selects the data provided by v i ,
• Commit: All consensus MECNs send commit message a request message to MECN M EC Nm will be sent and make
to others. After receiving the Prepare message, a replica a payment.
verifies and compares the verification result with the ⎧ pu ⎫
⎪
⎪ Kj ⎪
⎪
Pre-Prepare message from the MECN_leader. When a ⎨ ⎬
h(Blockt −1)
Prepare message is found to pass the feedback agreed DS_ Reqv j →M EC Nm : .
⎪
⎪ File Namei ⎪ ⎪
by 2 f replicas, it indicates that it enters the commit ⎩ ⎭
ti mestamp K v
state. The replicas broadcast a Commit message to other j
replicas. After the request has listened, the M EC Nm first verifies
• Reply (write block): Consensus MECNs write the result the certificate of v j and the existence of requested data
to blockchain. A replica will receive Commit information through IPFS. The M EC Nm then encrypts the data using
from other nodes. When it finds that the information has the requester’s public key and stores it as a single-access
been approved by 2 f + 1 (including itself), it considers link as an access authorization to the requester and secure
the transaction to reach a consensus and attempts to send protection of the data. When the data requester uses the link
the execution result to the block producer. to access the data to start the download, the link will be
During a period of time, each MECN receives a data storage invalid. Then, a smart contract sends payment to the data
request from different vehicles within communication range provider v i immediately, as shown in Fig. 7. The consensus
and generates a corresponding data record. These records will process for SCS is also adopted to PBFT, similar to SCR.
be gathered by the leader. And after the transaction is verified, Then the transaction record of the data sharing event will be
it is packaged into blocks and broadcasted to other pre-selected recorded on the blockchain, which is similar to data recording
MECNs on the blockchain for verification, which decide and data index recording.
whether to add the block to the end of the blockchain or not.
After the consensus is achieved, the data records contained
in the block will be persistent. More details about the con- C. Consensus Process for Blockchain
sensus process for blockchain are introduced in Section V. During a certain period, the leader will aggregate data
C. Other MECNs will save data synchronously through IPFS records, data index records, and data sharing records as well
network. as package them into blocks. This block will be added to the
(2) Data index recording: When v i uploads its data, it also end of the blockchain after the consensus is achieved. The con-
generates a corresponding data index for other vehicles to sensus process is adopted to BFT-DPoS consensus algorithm,
quickly filter data, as shown below: which can reduce the block generation interval to 0.5s and
make the block confirmation speed increase significantly [35].
Data i ndex i = (File Name||Descri pti on||
The DPoS is a consensus mechanism for authorizing equity.
Data Pr ovi der ||T i mestamp). Compared to the PoW and PoS mechanisms of the traditional
LI et al.: VEHICLE POSITION CORRECTION: A VEHICULAR BLOCKCHAIN NETWORKS-BASED GPS ERROR SHARING FRAMEWORK 907

TABLE III
DNN S ETTINGS

Fig. 7. Smart contract process for data sharing.

public blockchain, DPoS does not waste power resources to

compete for billing rights.
It votes to select a number of block producers, ensuring the
normal operation of the entire network. The MECNs on the
blockchain compete for block producers based on the number
of vehicle coins they hold, trust values, and computing power.
During the voting process, every MECN on the block uses the
rights in their hands to support trusted candidates and then
determines a certain number of producers based on the overall
voting situation, recorded as M EC N P (n), n = 1, 2, . . . , N.
Determine the production sequence for the M EC N P and take
turns to produce new blocks. Only one of the M EC N P is
allocated at a time for block production. Each M EC N P (n)
generates one block every 0.5 seconds, and a total of 6 Fig. 8. Comparison of positioning errors between different positioning
blocks are generated. If a producer misses producing block, methods.
it will be skipped and the producer will be deleted. Each time
a production cycle is completed, the producers will be re- the autonomous vehicle based on the comma.ai platform and
selected. Let the block producer become the primary node. an On-Board Unit (OBU) to obtain vehicle location, speed,
After the block is generated, the primary node broadcasts braking status and other parameters [49]. The speed of vehicle
the block to the remaining N − 1 nodes at the same time is 30 to 50km/ h. Vehicle acceleration is −4 to 4m/s 2 . The
and obtains the verification feedback from each node. If more travel direction of the vehicle is east or north. Each road
than 2/3 of the nodes agree with the verification, the block is section is approximately 1km. In particular, we analyze the
persistent and irreversible. location data to obtain the current GPS systematic error as well
as random error range and use it as simulation data. According
VI. S IMULATION R ESULTS AND D ISCUSSIONS
to the result of the error analysis, the systematic error E s is
In this section, the performance of our proposed algorithm about 20m and the random error obeys Gaussian distribution
is evaluated by a large number of simulations, which are Er ∼ N(0, 4.63). Detailed parameter settings for DNN are
divided into three aspects. First, we evaluated the accuracy listed in Table III.
and timeliness of the positioning error prediction algorithm. We divide the error evolution model sharing into the follow-
Next, the error correction accuracy of the proposed method ing three cases: 1) common vehicle i directly uses the position-
is evaluated by comparison with other positioning methods. ing error shared by sensor-rich vehicles to correct its position
Finally, by evaluating the robustness and security of the (case 1); 2) common vehicle i retrains DNN using data shared
algorithm, the impact of false data on the performance of by sensor-rich vehicles (case 2); 3) common vehicle i directly
CP methods is reflected. In the simulation, it is assumed use sensor-rich vehicles to share the prediction model to add
that all vehicles are equipped with the same GPS positioning a hidden layer and continue training (case 3). In addition,
performance. the sensor-rich vehicles that perform the positioning error
information sharing are set to 3. The threshold δ introduced
A. Positioning Error Prediction in Section IV.C. is set to 3m.
First, we verify the performance of the positioning error pre- For the sake of display, we extracted the first 80 samples.
diction algorithm in terms of accuracy and timeliness. We use It can be seen from Fig. 8, among these three cases,
908 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 22, NO. 2, FEBRUARY 2021

TABLE IV
E RROR C ORRECTION P ERFORMANCE E VALUATION OF D IFFERENT C ASES

the positioning error of case 1 is larger, while cases 2 and

3 have similar positioning performance. Although the fea-
sibility of the vehicle to improve the positioning accuracy
through positioning error sharing is obtained through analysis
in Section III, there are still differences between vehicles.
Therefore, when these error information are directly used
to improve the positioning accuracy of other vehicles, per- Fig. 9. Training curves tracking the learning loss of case 2 and case 3.
formance is affected. Cases 2 and 3 both re-estimate the
error evolution based on their own data and information
shared by sensor-rich vehicles, which is more suitable for
their own situation. In addition, we have listed some special
results in Table. IV, which represent the maximum, minimum,
average and mean square error (MSE) between the positioning
error and the true positioning error obtained in the three cases.
Similar to the results in Fig. 8, the difference between the
positioning error of case 1 and the true value is the largest,
with a maximum difference of 4.87 m. And the error between
case 2 and case 3 and the true value are within 2.5 m.
Similarly, it can be seen from the table that the MSE in each
case is 5.04, 1.24 and 0.53, respectively, indicating that the
prediction model in case 3 describes the experimental data
with better accuracy. However, the minimum column data
indicates that the prediction results in all three cases are very
similar to the real values, indicating that it is feasible to
improve the vehicle positioning accuracy by means of error
sharing.
We further evaluate the timeliness of case 2 and case 3 under Fig. 10. Comparison of positioning errors between different positioning
the same conditions. Learning loss of case2 and case3 are methods.
depicted in Fig. 9. It can be seen that learning loss of
error to some extent. Among these comparison algorithms,
case 3 drops faster than case 2, indicating that case 3 is
the worst performance is VCP, which has a maximum posi-
more efficient in learning. The reason is that for case 2,
tioning error of 9.56m. The underlying reason is that the
common vehicle i needs to retrain DNN based on data shared
positioning of the vehicle itself is not accurate, so there is an
by itself and sensor-rich vehicles. For case 3, the common
error in the position information provided during cooperative
vehicle i adds a layer of continuous training on top of the
positioning. In addition, when information is exchanged, V2V
other vehicle-trained network architecture. Since the previous
communication has delays, packet loss, etc., resulting in a
network model has been trained and the network parameters
decrease in positioning performance. By contrast, the position-
have been determined, the learning loss has to drop much
ing error of EMS and RRP is the smallest, because RSU and
faster, which is more time-efficient.
landmark have accurate position information. The maximum
positioning error of EMS and RRP are 4.15m and 6.57m.
B. Positioning Error Correction Although in more than half of the cases, the positioning error
Subsequently, we study the positioning performances of dif- of EMS, VCP, and RRP is within 3m. However, VCP and
ferent positioning methods, such as GPS, vehicular cooperative RRP have a larger error range than EMS. In addition, the RRP
positioning (VCP), RSU-based ranging positioning (RRP) and method requires at least two RSUs for positioning accuracy to
the proposed error model sharing (EMS) method. The results be guaranteed.
are shown in Fig. 10 and Table V. In this case, the GPS In order to further analyze the different positioning methods
maximum positioning error is close to 27m. It can be seen that to improve the GPS positioning error, we decompose the
EMS, VCP, and BRP all improve the accuracy of positioning positioning error into two directions: the east and the north.
LI et al.: VEHICLE POSITION CORRECTION: A VEHICULAR BLOCKCHAIN NETWORKS-BASED GPS ERROR SHARING FRAMEWORK 909

Fig. 12. Comparison of positioning errors between different positioning

methods with fewer cooperators.

TABLE VI
ROBUSTNESS AND S ECURITY E VALUATION OF D IFFERENT M ETHODS

error provides the same error magnitude and direction for

each GPS receiver within a certain range, as we analyzed in
Fig. 11. Positioning Error of different methods. Section III. The VCP method does not eliminate the systematic
error. There is still a large positioning error. Both EMS and
TABLE V RRP methods can better eliminate GPS systematic errors,
A CCURACY E VALUATION OF D IFFERENT M ETHODS resulting in higher positioning accuracy.

C. Robustness and Security Analysis

In order to evaluate the robustness of the proposed method,
we set the vehicles for cooperative positioning in the VCP
and RSUs in RRP to 2. For the EMS method, set the
vehicles for data sharing to 2 and 1, respectively. The VCP
method and RRP method cannot perform positioning error
correction if only one vehicle or RSU available. The results
As can be seen from the Fig. 11 and Table V, the positioning are depicted in Fig. 12 and Table VI. EMS (2) indicates
errors of the EMS and RRP methods in both directions are the use of information shared by two sensor-rich vehicles,
shifted in the positive and negative directions with y = 0 as while EMS (1) indicates the use of one. Combined with
the center, while the error of the VCP method is not centered Fig. 10, the performance of the error correction for each
on y = 0. In addition, the maximum positioning errors of method may decrease as the number of vehicles or RSUs used
EMS, VCP, and RRP in a single direction are 3.58m (north), for cooperation is reduced. For VCP, when the cooperative
8.07m (east) and 6.00m (east), respectively. The result verifies positioning vehicle changes from 3 to 2, the maximum posi-
that the vehicle GPS positioning error can be decomposed tioning error increases by 10.89m. Similarly, the maximum
into systematic error and random error, wherein the systematic positioning error of RRP is increased by 3.66m. In contrast,
910 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 22, NO. 2, FEBRUARY 2021

and efficiently improving the GPS positioning accuracy of

vehicles. In particular, by analyzing the GPS positioning error
characteristics, the feasibility of using sensor-rich vehicles to
compensate for the positioning accuracy of common vehicles
is obtained. Next, for the landmark discontinuous scenario,
the DNN-based error correction algorithm has been proposed
to obtain the positioning error evolution. Then, we have
designed a vehicular blockchain based on vDLT platform for
secure and efficient information storage and sharing. In addi-
tion, by fully exploiting the characteristics of blockchain and
edge computing, smart contracts for data recording and sharing
have been developed, which are running on the vehicular
blockchain. In the proposed framework, the MECNs not
only act as storage units, but also perform smart contracts
and perform DNN operations. Extensive simulations have
exhibited the effectiveness and accuracy of our proposed
method compared with the traditional method. For many error
prediction models stored on blockchain, how to choose the best
provider to achieve higher positioning accuracy is one of the
Fig. 13. Comparison of positioning errors between different positioning research priorities of future work. In addition, considering the
methods with malicious node.
impact of multipath in urban environments, how to eliminate
random errors is another problem that needs to be studied.
EMS still maintains good performance. When only two sensor-
rich vehicle data are available, the maximum positioning R EFERENCES
error increases from 4.15m to 6.79m. In addition, even if [1] S. Demetriou, P. Jain, and K. Kimx, “CoDrive: Improving automobile
positioning via collaborative driving,” in Proc. Int. Conf. IEEE Comput.
only one sensor-rich vehicle data is available, the positioning Commun. (IEEE INFOCOM), Apr. 2018, pp. 1–9.
error correction performance is still higher than VCP (three [2] L. Wei, C. Cappelle, and Y. R. Wei, “Camera/laser/GPS fusion method
cooperative vehicles) and RRP (two RSUs). It shows that for vehicle positioning under extended NIS-based sensor validation,”
IEEE Trans. Instrum. Meas., vol. 62, no. 11, pp. 3110–3122, Nov. 2013.
the EMS method is more adaptable to the environment, and
[3] M. Sefati, M. Daum, B. Sondermann, K. D. Kreiskother, and
can still achieve acceptable error correction results without A. Kampker, “Improving vehicle localization using semantic and pole-
sufficient sensor-rich vehicles. like landmarks,” in Proc. Int. Conf. Intell. Veh. Symp. (IV), 2017,
Next, for a situation where there are three vehicles or three doi: 10.1109/IVS.2017.7995692.
[4] G. De Angelis, G. Baruffa, and S. Cacopardi, “GNSS/cellular hybrid
RSUs co-located, we randomly set one of them as a malicious positioning system for mobile users in urban scenarios,” IEEE Trans.
node. Malicious nodes provide false location information, Intell. Transp. Syst., vol. 14, no. 1, pp. 313–321, Mar. 2013.
which affects the correction of positioning errors. We evaluate [5] N. Alam and A. G. Dempster, “Cooperative positioning for vehicular
networks: Facts and future,” IEEE Trans. Intell. Transp. Syst., vol. 14,
the performance of the three positioning methods in this case, no. 4, pp. 1708–1717, Dec. 2013.
and the results are shown in Fig. 13 and Table VI. Case (a) [6] M. Yassin and E. Rachid, “A survey of positioning techniques and
represents that the false data provided by the malicious node is location based services in wireless networks,” in Proc. IEEE Int. Conf.
Signal Process. Informat. Commun. Energy Syst., Feb. 2015, pp. 1–5.
slightly different from the real data (i.e., positioning error < δ), [7] M. Ke, Y. Xu, A. Anpalagan, D. Liu, and Y. Zhang, “Distributed
and case (b) represents that the data error is large. First, TOA-based positioning in wireless sensor networks: A potential game
false data has a large impact on the performance of VCP and approach,” IEEE Commun. Lett., vol. 22, no. 2, pp. 316–319, Feb. 2018.
RRP, and the positioning error increases with the increase of [8] V. Pasku, A. De Angelis, M. Dionigi, A. Moschitta, G. De Angelis, and
P. Carbone, “Analysis of nonideal effects and performance in magnetic
false data error. For the EMS algorithm, when the false data positioning systems,” IEEE Trans. Instrum. Meas., vol. 65, no. 12,
difference is small, the common vehicle will still accept the pp. 2816–2827, Dec. 2016.
positioning error correction, thus increasing the positioning [9] Y. Wang, X. Duan, D. Tian, M. Chen, and X. Zhang, “A DSRC-based
vehicular positioning enhancement using a distributed multiple-model
error. When the false data difference is large, the data will Kalman filter,” IEEE Access, vol. 4, pp. 8338–8350, 2016.
be eliminated without affecting the final positioning accuracy. [10] J. Li, J. Gao, H. Zhang, and T. Z. Qiu, “RSE-assisted lane-level position-
In addition, due to the nature of the blockchain, the data ing method for a connected vehicle environment,” IEEE Trans. Intell.
Transp. Syst., vol. 20, no. 7, pp. 2644–2656, Jul. 2019, doi: 10.1109/
written to the blockchain cannot be tampered with. There- TITS.2018.2870713.
fore, for the false data provider, it can be directly removed [11] O. Walter, J. Schmalenstroeer, A. Engler, and R. Haeb-Umbach,
from the blockchain system, thus ensuring the accuracy and “Smartphone-based sensor fusion for improved vehicular navigation,”
in Proc. 10th IEEE Workshop Positioning Navigat. Commun., Mar. 2013,
security of the data. Therefore, the positioning error correction pp. 1–6.
performance is similar to that of EMS (2) in Fig. 12. [12] J. Kang et al., “Blockchain for secure and efficient data sharing in
vehicular edge computing and networks,” IEEE Internet Things J., vol. 6,
no. 3, pp. 4660–4670, Jun. 2019, doi: 10.1109/JIOT.2018.2875542.
VII. C ONCLUSION [13] C. Qiu, F. R. Yu, H. Yao, C. Jiang, F. Xu, and C. Zhao, “Blockchain-
based software-defined industrial Internet of Things: A dueling deep
In this paper, a vehicular blockchain-based positioning Q-learning approach,” IEEE Internet Things J., vol. 6, no. 3,
error correction framework has been proposed for securely pp. 4627–4639, Jun. 2019, doi: 10.1109/JIOT.2018.2871394.
LI et al.: VEHICLE POSITION CORRECTION: A VEHICULAR BLOCKCHAIN NETWORKS-BASED GPS ERROR SHARING FRAMEWORK 911

[14] J. Xie et al., “A survey of blockchain technology applied to smart [37] B. Leiding, P. Memarmoshrefi, and D. Hogrefe, “Self-managed and
cities: Research issues and challenges,” IEEE Commun. Surveys Tuts., blockchain-based vehicular ad-hoc networks,” in Proc. Int. Conf. ACM
vol. 21, no. 3, pp. 2794–2830, 3rd Quart., 2019, doi: 10.1109/COMST. Pervasive Ubiquitous Comput., Adjunct, Sep. 2016, pp. 137–140.
2019.2899617. [38] A. Dorri, M. Steger, S. S. Kanhere, and R. Jurdak, “Blockchain: A
[15] F. P. Müller, “Survey on ranging sensors and cooperative techniques for distributed solution to automotive security and privacy,” IEEE Commun.
relative positioning of vehicles,” Sensors, vol. 17, no. 271, pp. 1–27, Mag., vol. 55, no. 12, pp. 119–125, Dec. 2017.
Feb. 2017, doi: 10.3390/s17020271. [39] M. Singh and S. Kim, “Blockchain based intelligent vehicle data
[16] S. Kuutti, S. Fallah, K. Katsaros, M. Dianati, F. Mccullough, and sharing framework,” Jul. 2017, arXiv:1708.09721. [Online]. Available:
A. Mouzakitis, “A survey of the state-of-the-art localization techniques https://arxiv.org/abs/1708.09721
and their potentials for autonomous vehicle applications,” IEEE Inter- [40] Z. Yang, K. Yang, L. Lei, K. Zheng, and V. C. M. Leung, “Blockchain-
net Things J., vol. 5, no. 2, pp. 829–846, Apr. 2018, doi: 10.1109/ based decentralized trust management in vehicular networks,” IEEE
JIOT.2018.2812300. Internet Things J., vol. 6, no. 2, pp. 1495–1505, Apr. 2019, doi: 10.1109/
[17] K.-W. Chen et al., “Vision-based positioning for Internet-of-vehicles,” JIOT.2018.2836144.
IEEE Trans. Intell. Transp. Syst., vol. 18, no. 2, pp. 364–376, [41] Y. Fu, C. Li, H. Luan, Y. Zhang, and G. Mao, “Infrastructure-cooperative
Feb. 2017. algorithm for effective intersection collision avoidance,” Transport.
[18] E. Javanmardi, Y. Gu, M. Javanmardi, and S. Kamijo, “Autonomous Res. C, Emer., vol. 2018, no. 89, pp. 188–204, 2018.
vehicle self-localization based on abstract map and multi-channel [42] S. Miura, L. T. Hsu, F. Chen, and S. Kamijo, “GPS error correction with
LiDAR in urban area,” IATSS Res., vol. 43, no. 1, pp. 1–13, Apr. 2019. pseudorange evaluation using three-dimensional maps,” IEEE Trans.
[19] A. Y. Hata and D. F. Wolf, “Feature detection for vehicle localization Intell. Transp. Syst., vol. 16, no. 6, pp. 3104–3115, Dec. 2015.
in urban environments using a multilayer LIDAR,” IEEE Trans. Intell. [43] I. Skog and P. Hndel, “In-car positioning and navigation technologies–
Transp. Syst., vol. 17, no. 2, pp. 420–429, Feb. 2016. A survey,” IEEE Trans. Intell. Transp. Syst., vol. 10, no. 1, pp. 4–21,
[20] J. K. Suhr, J. Jang, D. Min, and H. G. Jung, “Sensor fusion-based low- Mar. 2009.
cost vehicle localization system for complex urban environments,” IEEE [44] J. Du and M. J. Barth, “Next-generation automated vehicle location
Trans. Intell. Transp. Syst., vol. 18, no. 5, pp. 1078–1086, May 2017. systems: Positioning at the lane level,” IEEE Trans. Intell. Transp. Syst.,
[21] H. Kloeden, D. Schwarz, E. M. Biebl, and R. H. Rasshofer, “Vehicle vol. 9, no. 1, pp. 48–57, Mar. 2008.
localization using cooperative RF-based landmarks,” in Proc. Int. Conf. [45] A. Courville, I. Goodfellow, Y. Bengio, Deep Learning. Cambridge, MA,
Intell. Veh. Symp. (IV), Jul. 2011, pp. 387–392. USA: MIT Press, 2016.
[22] S. Venkatraman, J. Caffery, and H. R. You, “A novel ToA location [46] M. T. Hagan, H. B. Demuth, M. H. Beale, and O. De Jesús, Neural
algorithm using LoS range estimation for NLoS environments,” IEEE Network Design. Cambridge, MA, USA: Martin Hagan, 2014.
Trans. Veh. Technol., vol. 53, no. 5, pp. 1515–1524, Sep. 2004. [47] A. Ng, “Machine learning,” Stanford Univ., Stanford, CA, USA,
[23] N. Drawil and O. Basir, “Intervehicle-communication-assisted localiza- Jul. 2008. [Online]. Available: https://www.youtube.com/view_play_list?
tion,” IEEE Trans. Intell. Transp. Syst., vol. 11, no. 3, pp. 678–691, p=A89DCFA6ADACE599
Sep. 2010. [48] (Feb. 9, 2018). EOS Storage Whitepaper. [Online]. Available:
[24] G. M. Hoang, B. Denis, J. Härri, and D. T. M. Slock, “Distributed https://medium.com/eosio/start-an-eos-meetup-in-your-community-
link selection and data fusion for cooperative positioning in GPS-aided 761be355fce
IEEE 802.11p VANETs,” in Proc. IEEE Workshop Positioning Navigat. [49] E. Santana and G. Hotz, “Learning a driving simulator,” Aug. 2016,
Commun., Mar. 2015, pp. 1–6. arXiv:1608.01230. [Online]. Available: https://arxiv.org/abs/1608.01230
[25] A. Kakkavas, M. H. C. Garcia, R. A. Stirling-Gallacher, and
J. A. Nossek, “Multi-Array 5G V2V relative positioning: Performance
bounds,” in Proc. Int. Conf. IEEE Global Commun. (GLOBECOM), Abu
Dhabi, United Arab Emirates, Dec. 2018, pp. 206–212.
[26] H. Aly, A. Basalamah, and M. Youssef, “Robust and ubiquitous
smartphone-based lane detection,” Pervasive Mobile Comput., vol. 26,
pp. 35–56, Feb. 2016.
[27] J. Yin, Q. Wan, S. Yang, and K. C. Ho, “A simple and accurate TDOA-
Changle Li (M’09–SM’16) received the Ph.D.
AOA localization method using two stations,” IEEE Signal Process.
degree in communication and information system
Lett., vol. 23, no. 1, pp. 144–148, Dec. 2016.
from Xidian University, China, in 2005. He con-
[28] S. Bozkurt, G. Elibol, S. Gunal, and U. Yayan, “A comparative study on
ducted his Post-Doctoral Research in Canada and
machine learning algorithms for indoor positioning,” in Proc. Int. Symp.
the National Institute of information and Commu-
Innov. Intell. Syst. Appl. (INISTA), Madrid, Spain, 2015, pp. 1–8.
nications Technology, Japan, respectively. He had
[29] M. Becker and B. Ahuja, “Implementing real-life indoor positioning been a Visiting Scholar with the University of Tech-
systems using machine learning approaches,” in Proc. 8th Int. Conf. nology Sydney. He is currently a Professor with
Inf., Intell., Syst. Appl. (IISA), Larnaca, Cyprus, 2017, pp. 1–6. the State Key Laboratory of Integrated Services
[30] Z. Zheng, S. Xie, H. Dai, X. Chen, and H. Wang, “An overview Networks, Xidian University. His research interests
of blockchain technology: Architecture, consensus, and future trends,” include intelligent transportation systems, vehicular
in Proc. IEEE BigData Congr., Honolulu, HI, USA, Jun. 2017, networks, mobile ad hoc networks, and wireless sensor networks.
pp. 557–564.
[31] S. Nakamoto. (2008). Bitcoin: A Peer-to-Peer Electronic Cash System.
[Online]. Available: http://www.ildgroup.si/uploads/product/20/bitcoin.
pdf
[32] (Dec. 2017). Ethereum. [Online]. Available: https://www.ethereum.org/
[33] D. Larimer. (Nov. 2013). Transactions as Proof-of-Stake.
[Online]. Available: https://bravenewcoin.com/assets/Uploads/
TransactionsAsProofOfStake10.pdf
Yuchuan Fu received the B.Eng. degree in telecom-
[34] M. Castro and B. Liskov, “Practical byzantine fault tolerance and munication engineering from the Xi’an University of
proactive recovery,” ACM Trans. Comput. Syst., vol. 20, no. 4, Posts and Telecommunications, China, in 2014. She
pp. 398–461, Nov. 2002.
is currently pursuing the Ph.D. degree in telecom-
[35] (Mar. 16, 2018). EOS.IO Technical White Paper v2. [Online]. munication engineering with Xidian University,
Available: https://github.com/EOSIO/Documentation/blob/master/ Xi’an, China. Her current research interests include
TechnicalWhitePaper.md#consensus-algorithm-bft-dpos communication protocol and algorithm design in
[36] A. Kosba, A. Miller, E. Shi, Z. Wen, and C. Papamanthou, “Hawk: vehicular networks and wireless sensor networks.
The blockchain model of cryptography and privacy-preserving smart
contracts,” in Proc. IEEE SP, San Jose, CA, USA, May 2016,
pp. 839–858.
912 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 22, NO. 2, FEBRUARY 2021

Fei Richard Yu (S’00–M’04–SM’08–F’18) Tom H. Luan (M’14–SM’18) received the B.Eng.

received the Ph.D. degree in electrical engineering degree from Xi’an Jiaotong University, China,
from The University of British Columbia in 2003. in 2004, the M.Phil. degree from The Hong Kong
From 2002 to 2006, he was with Ericsson, Lund, University of Science and Technology in 2007, and
Sweden, and a start-up in California, USA. He joined the Ph.D. degree from the University of Waterloo,
Carleton University in 2007, where he is currently Ontario, Canada, in 2012. He is currently a Professor
a Professor. His research interests include wireless with the School of Cyber Engineering, Xidian Uni-
cyber-physical systems, connected/autonomous versity, Xi’an, China. He has authored/coauthored
vehicles, security, distributed ledger technology, more than 40 journal articles and 30 technical arti-
and deep learning. He is a registered Professional cles in conference proceedings. He awarded one
Engineer in ON, Canada. He is a fellow of the U.S. patent. His research mainly focuses on content
Institution of Engineering and Technology. He is a Distinguished Lecturer, distribution and media streaming in vehicular ad hoc networks and peer-
the Vice President (Membership), and an Elected Member of the Board of to-peer networking and the protocol design and performance evaluation of
Governors of the IEEE Vehicular Technology Society. He received the IEEE wireless cloud computing and edge computing. He served as a TPC Member
Outstanding Service Award in 2016, the IEEE Outstanding Leadership Award for IEEE Globecom, ICC, and PIMRC; and the Technical Reviewer for
in 2013, the Carleton Research Achievement Award in 2012, the Ontario multiple IEEE Transactions, including TMC, TPDS, TVT, TWC, and ITS.
Early Researcher Award (formerly Premiers Research Excellence Award)
in 2011, the Excellent Contribution Award from IEEE/IFIP TrustCom
2010, the Leadership Opportunity Fund Award from the Canada Foundation Yao Zhang received the B.Eng. degree in telecom-
of Innovation in 2009, and the Best Paper Awards from IEEE VTC munication engineering from the Xi’an University
2017 Spring, ICC 2014, Globecom 2012, IEEE/IFIP TrustCom 2009, and of Science and Technology, China, in 2015. He is
International Conference on Networking 2005. He has served as the technical currently pursuing the Ph.D. degree in telecommu-
program committee co-chair for numerous conferences. He serves on the nication engineering with Xidian University, Xi’an,
Editorial Boards for several journals, including the Co-Editor-in-Chief China. His current research interests include com-
for Ad Hoc and Sensor Wireless Networks, a Lead Series Editor for munication protocol and performance evaluation of
the IEEE T RANSACTIONS ON V EHICULAR T ECHNOLOGY, the IEEE vehicular networks, edge caching, and wireless sen-
T RANSACTIONS ON G REEN C OMMUNICATIONS AND N ETWORKING, and sor networks.
the IEEE C OMMUNICATIONS S URVEYS AND T UTORIALS .