Bimodal Inertial Sensor-Based Gait Analysis Using Orientation And Time-Invariant Features
Wearable sensors technology has enabled continuous monitoring of human gait to be used as behavioural biometric authentication. A system that authenticates the identity of an individual from his/her gait signature is known as gait recognition. This thesis focuses on gait analysis that extracts infor...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2020
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Wearable sensors technology has enabled continuous monitoring of human gait to be used as behavioural biometric authentication. A system that authenticates the identity of an individual from his/her gait signature is known as gait recognition. This thesis focuses on gait analysis that extracts information from inertial sensors (accelerometer and gyroscope) embedded in smartphones for identity recognition. Gait signals captured by inertial sensors are shown to be reliable and have promising use because they are unobtrusive and user friendly. However, it is challenging to evaluate the impact of multiple devices at different on-body positions and it also has a great impact on system accuracy. Besides, time lapse with respective covariate factors (i.e. clothing, shoes, etc.) has also been reported to cause a significant drop in recognition performance. Recent research has been conducted to investigate these problems but the data used in the studies were laboratory-controlled. Either covariates like clothing, footwear, environment and walking speed were controlled or the sensor position was fixed on the participant’s body. This is unlikely to happen in real life so the good results reported might be unreliable. In this thesis, methods have been proposed to solve the pose variation and time lapse problems. Firstly, a Gait Inertial Gaussian Kernels with Randomized Kernel Extreme Learning Machine (GIGRKELM) method is introduced to address the effect of different device-instances located at different on-body position for gait recognition. Secondly, an Adaptive 1-Dimensional Local Domain Adaptation (A1LDA) is introduced to solve time-variation problem with the respective covariate factors. The proposed methods have shown promising results demonstrating that gait signals extracted from inertial sensors can be used as a reliable means of biometrics. On top of that, a self-collected dataset called MMUISD is presented in this thesis. The dataset is collected to be as close to the real world as possible. The first and second data collection sessions are separated by a 1-month time gap with 330 and 144 participants, respectively. The dataset contains rich variations that include: (1) signals from different types of android phones, (2) different sensor placements without fixing the orientation/positions, and (3) three different walking speeds. This is the first publicly available inertial sensor dataset containing time-variation. |
|---|