Use this URL to cite or link to this record in EThOS:
Title: View-invariant feature selector and its application on gait recognition
Author: Jia, Ning
ISNI:       0000 0004 6495 7549
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2016
Availability of Full Text:
Access from EThOS:
Access from Institution:
The proliferation of the national-wide deployment of surveillance cameras and identity management systems has promoted the development of biometric systems. Gait as a behavioural biometric trait can be measured unobtrusively at a moderate distance, thus it is predominant in remote human tracking and identification tasks. The past two decades have witnessed a considerable development of gait recognition systems. Yet there are challenges that confine the practical application of gait analysis. The motivation of our work is to identify the problems and find corresponding solutions to explore the potentials of gait recognition and promote its applicability in open-world scenarios. Gait recognition systems use human profile as features, while the appearance of human profile, also known as silhouette, can be affected in various manners. For example, clothing changes the shape of torso (coat) or legs (skirt); carrying bag attaches extra region to the silhouette; walking surface or speed variation changes the appearance of legs. On the other hand, camera viewpoint variation changes the shape of both the upper and lower body, while segmentation errors may cause massive corruption of the gait features. We summarise them into two categories: partial interference and holistic deformation. The former has been well addressed by existing literatures. The holistic deformation on gait silhouette results in large intra-class variation, and we notice that the performance of conventional approaches decreases under such circumstance. Thus our work focus mostly on the latter challenge. Accordingly, we propose ViFS, an automatic feature selection approach that seeks for the optimal representation features from gallery set, and evaluate its performance under various conditions. We find that ViFS minimises the intra-class variation between gallery and probe data, and by introducing proper feature enhancers, we can further reduce the number of holistic deformation modalities required in the gallery set. We test the proposed method on public dataset that contains viewpoint variations, and the matching accuracy has achieved 99.1% on CASIA Dataset B and 97.7% on OU-ISIR Large Population Dataset. The formulation and discussion are presented in Chapter 3. The success of Convolutional Neural Network (CNN) based methods in image classification field has drawn attention from researchers. Recently a large number of literatures have covered the application of CNN in computer vision tasks, including face and gait recognition in the biometrics field. CNN has much greater discriminant learning ability in the highly non-linear space. Thus we merge CNN feature maps with the proposed ViFS approach, which achieves the state-of-the-art performance on view-invariant gait recognition problem. The methodology and results are presented in Chapter 4. Among the holistic deformation challenge, the silhouette quality issue is seldom addressed, while no published dataset concerns with the influence of segmentation quality on gait recognition algorithms. We create a dataset that contains silhouettes with six different segmentation qualities in both gallery and probe set, and evaluated the conventional methods as well as the proposed ViFS approach on this dataset. It is proved that ViFS based framework and its extension outperforms the conventional methods by 8%-10%, which further indicates the effectiveness of ViFS based framework on gait holistic deformation challenge. This work is presented in Chapter 5. This thesis aims at tackling the gait silhouette holistic deformation challenge, and ViFS based frameworks are proposed to achieve robust recognition performance. We evaluate the effect of different feature enhancers for ViFS, and find out that the discriminant power of CNN feature maps is much more powerful than subspace learning methods (3% higher accuracy under same conditions), thus it requires less gallery data to achieve deformation-invariant recognition.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: TK Electrical engineering. Electronics Nuclear engineering