Human Identification in Video Surveillance Using Improved Gait Recognition Approaches

Research on several biometrics is on going with the aim to achieve robust human identification in visual surveillance environments. Compared to other biometrics, gait has attracted significant attention in recent years because of its unique advantages, which other biometrics may not offer. Mainly, it can be used with low–resolution video feeds, obtained at a distance, without alerting the subject. The most prominent example of using gait in security access, person verification/identification application scenario is shown in Figure 1.

Interest in gait as a biometric has increased significantly due to the promising recognition results obtained from research in this area, under constrained environments. Recent research has focused more on improving recognition results in realistic environments, where it is necessary to address the effects of changes in view, illumination changes, and fluctuation of gait patterns due to the subject carrying goods, changes in footwear, cloths and the walking surface. Methods proposed to date have struggled to provide good gait recognition performance in these realistic environments.

The broad scope of this PhD research program is to develop improved gait recognition techniques that can perform well in realistic unconstrained environments. Within this broadscope, the following main tasks are being carried out:

1. Gait recognition from reconstructed 3D data obtained from multiple views in a multi-camera surveillance environment.
   -Gait energy volume (GEV) is proposed as novel 3D gait feature that represent 3D gait temporal information in a single volume.
2. Frontal view gait recognition that enables the usage of gait with other biometrics in the application areas such as portal-based human identification.
   -Frontal volumes are extracted from frontal depth data and frontal gait energy volumes are initially investigated.
3. Explore view invariant and cross-capture modality features:
   -Backfill gait energy images (BGEI) are proposed and initial experiments shows state-of-the art recognition results.
4. Model based techniques in 2D and 3D
   -Ellipsoidal based modal fitting and Fourier based gait feature extraction are investigated and 3D model fitting based on annealed particle filters are being explored.
5. Feature modelling techniques and optimised classifier for the high-performing gait recognition
   -Local directional pattern (LDP) based feature modelling technique is proposed and sparse representation based classifier (SRC) is chosen as best performing classifier in gait context.
6. Collect an in-house gait database to facilitate gait recognition research in several challenging factors.
   -Frontal depth gait data is collected for 35 subjects with 6 different walking conditions. Examples are shown in Figure 2