Human Activity Recognition (HAR) systems aim to recognize human activities through sensors in order to provide assistance. The key steps in designing a HAR system are:
1) Acquiring sensor data and preprocessing it by removing noise.
2) Segmenting the preprocessed data into windows that may contain activities.
3) Extracting features from each window to reduce the data into discriminative features.
4) Training a classification model on the extracted features to predict activity labels, and evaluating the model's performance using methods like a confusion matrix.
2. Introduction
The goal of HAR is to provide information on a user’s
behaviour that allows computing systems to
proactively assist users with their tasks.
Why do we need HAR?
10. How would you go about designing a HAR
System for you project?
● ARC (Activity Recognition Chain)
○ Sensor Data Acquisition
○ Preprocessing
○ Data Segmentation
○ Feature Extraction and Selection
○ Training + Classification
○ Evaluation
● Challenges you might face
15. Sensor Data Acquisition 1/6
Sensor Output Format
Si
= (d1
, d2
,... dj
, …., dt
)
i : ith
sensor | j : output of sensor at time t = j
● dj
’s can be multidimensional.
● Each sensor is sampled at regular intervals. However, sampling rates
of different types of sensors might differ.
18. Preprocessing 2/6
Sensor data Preprocessing Preprocessed Time Series D/
(synchronization, noise removal)
n: total data dimensions, t: # of samples.
● Generic, only dependent on data.
● Algorithms which preserve relevant
signal characteristics, only filter noise.
19. How can she make her data usable to extract
features?
20. Data Segmentation 3/6
To identify time-based segments of preprocessed data that may contain
relevant information.
D/
W = { w1
, w2
, …, wi
, …, wm
} | wi
= (tstart
, tend
)
Different Techniques
● Sliding Window
● Energy Based
● Additional Sensors, Contextual Sources
22. Feature Extraction & Selection 4/6
Reducing signals into features that are discriminative of the activity.
Feature Vector Xi
= F(D/
,wi
) | Xi
: [ ]n x 1
where n: # of features
Features
● Corresponding to the same activity should be clustered together.
● Can be based on
○ Signals, Body-Model
● There are built-in methods for feature ranking, selection.
23. More features More dimensions in data
Trade-offs: memory, computational power
24. How can she build a model using the data which
classifies live data streams into activities?
25. Training 5/6
Training Data T = { (Xi
, yi
) }i = 1 to N
| Xi
: Feature Vector, yi
: Label
Goal : To minimize classification error on T.
Training Method
● an algorithm which trains the model to predict label yi
, given a
feature vector Xi
● Selected on the basis of
○ Type of Activity
○ Complexity of Feature Space
26. Classification 5.5/6
A two-step process
1. Each Xi
is assigned confidence values w.r.t class labels yi
’s.
2. The class label with the highest confidence value is chosen.
If all confidence values are below a certain threshold, a null class label is
assigned to the corresponding data sample.