Recent interest in gait as a biometric

Get Complete Project Material File(s) Now! »

Image capturing

The gait motion was captured using a Sony SNC-RZ30P network camera. Capture software for the system was written to facilitate the acquisition of unprocessed video data from the camera. The software recorded video footage contained the subjects as well as the background before the capture of each subject for the process of background estimation and segmentation. The recorded video frames were streamed as unprocessed raw JPG (320×240) data back to the host computer at 25 frames per second and then saved in local memory.
Participation in the gait data collection was voluntary. Prior to the collection process the subject were asked to read, understand and sign an approved consent form. The participation consent form and information sheet is shown in Appendix A.1. The subjects participated were not required to wear any special type of clothing or footwear. The aim was to capture the subjects gait motion in their everyday wear of clothing and footwear. Prior to the recording, the subjects were given time to accustom themselves with walking around the pathway. The subjects were asked to enter the camera view field by placing their right leg in front on the marked position.
For each subject the following gait motion was captured:
– six gait sequences from right-to-left camera view field
– six gait sequences from left-to-right camera view field
– slow pace gait capture
– fast pace gait capture
For each gait sequence, at least 2 complete walk cycles of gait motion for fast pace and 3 walk cycles for slow pace were captured. The database compromises of 5 female and 20 male subjects. The age of subjects varied from 18 to 55 years. The height of the subjects ranged from 160cm to 186cm and the weight ranged from 55Kg to 110Kg. A full list can be found in Appendix A.3. Figure 3-8 shows an image capture of a subject walking from right-to-left of the camera view field. With the backlight setting, the subject being recorded appears to be darker compared to the brightly light background.

Use of a cheap camera and its limitation

For the purpose of comparison and practicality, experiments were also conducted with an alternate cheaper camera; 4XEM WPT camera. The camera had features such as Pan, Tilt and digital zoom with board lens with 4.3mm focal length. The camera was placed in the same gait setup as Sony camera. The gait images were acquired through the VS-IP surveillance software. Given the restriction of the gait lab space in controlled environment, the gait image  acquired using the camera had a major shortfall of limited number of gait cycles. On average only one and half walk-cycle of gait sequence could be achieved per subject. Ideally two walk cycles was required to be able to explore the number of gait cycles required for gait recognition and for comparison purposes. The results of gait recognition using 4XEM camera is shown in Appendix D. Furthermore the camera lens produced images with fish-eye effect. This somewhat distorted the images which could have resulted in variation of gait features. This could be attributed to the lower results achieved when compared to results of Sony camera.

READ  Railway system and passenger transportation planning 

Equal Error Rate Analysis

As discussed in Section 4.6, the performance of the proposed system was also measured in
terms of equal-error rate (EER).

EER analysis for view-dependent experiment

For each view angle, EER analysis based on REI+CEP and GEI+CEP was also conducted based on the experiments conducted in the above sections. Table 5-8 shows the EER analysis results. An average EER of 4.12% is achieved for REI+CEP and 6.38% for GEI+CEP in normal pace condition. This could be attributed to the matching score distribution. Figure 5-12 shows the probability density function plot for normal pace side-view matching scores. As shown in the graph, GEI based cepstral coefficients approach has a much narrower score distribution of the correct and not-correct matching score, hence this increases the overlapping area of the two distribution, resulting in much higher EER. The REI based cepstral coefficients matching score distribution has a much wider score distribution of correct and not-correct matching score, hence the area of overlapping between the two score distribution is much smaller, resulting in better EER. Results indicate that REI based features based system performs better in differentiating between correct and non-correct matches.

1 Introduction
1.1 Biometrics
1.2 Gait as a Behavioural Biometric
1.2.1 Principles of gait
1.2.2 Recent interest in gait as a biometric
1.3 Research Contribution
1.4 Thesis Outline
2 Background and Related Work .
2.1 Introduction .
2.2 Model-based Approaches .
2.3 Holistic-based Approaches
2.4 Gait Datasets
2.5 Conclusion
3 Image Acquisition and Pre-processing
3.1 Introduction
3.1.1 Assumptions
3.1.2 Overview of the method
3.2 Initial Gait Lab Setup
3.3 Image Acquisition
3.3.1 Lab set-up
3.3.2 Image capturing
3.4 Background Modelling and Segmentation.
3.5 Gait Cycle Detection
3.6 Conclusion
4 Sub-Vector Quantisation 
4.1 Overview
4.2 Gait Feature
4.2.1 Height normalisation
4.2.2 Width vectors.
4.2.3 Sub-vector gait feature representation
4.3 LBG-VQ Algorithm
4.3.1 Vector quantisation
4.3.2 Sub-vector segment and codebook size
4.4 Gait Sequence Matching .
4.4.1 Dynamic time warping
4.4.2 Classification
4.5 Data
4.5.1 CMU gait database
4.5.2 CASIA-B gait database
4.5.3 SOTON small gait database
4.5.4 AUCKLAND gait database
4.6 Experimental Results .
4.6.1 Results for CMU gait database
4.6.2 Results for CASIA-B gait database1
5 Cepstral Feature Based Gait Recognition 
6 Multi-view Gait Recognition Using Sparse Representation
7 Conclusions and Future Work 
Bibliography
Appendices

GET THE COMPLETE PROJECT
Modelling Gait Attributes Using Single and Multiple Viewpoints for Human Identification

Related Posts