On the analysis of fronto-normal gait for human recognition

The objective of this thesis is to investigate techniques which will enhance the use of biometrics. Current events show that the threat of terrorism is very real. Monitoring surveillance cameras full time would require tremendous human resources. The availability of low cost, high powered computers and video cameras allows us to apply computer vision techniques to help in identifying suspicious personnel. We investigate approaches to human recognition that are practical and robust. Currently, many biometrics are intrusive and may require that a subject be near a camera in order to work. Other biometrics can be measured from afar. One of these is gait, which has been the subject of recent research. Since gait involves movement, time is implicit in its measurement. The spatio-temporal characteristics of gait allows the vast array of Time Series Analysis methods to be used to describe gait. Currently, gait research has been done on a frontal parallel walk, which requires a clear field of view over a large area and uses standard measures of gait. However, a common scenario is where people have to queue up to access a facility. Because of physical constraints, they walk towards a camera which may be placed at the point of access. This camera captures the fronto-normal view of a user. The physical space needed is reduced and we can employ a far-distance biometric like gait. This approach is thus practical furthermore, this view of humans is also robust, as we can combine gait with a near-distance biometric as face in a natural way. In this thesis, we explore the usefulness of fronto-normal gait by tracking body parts as compared to silhouette based methods as it gives more information. The main contributions to existing knowledge are as follows: First is the use of advanced Time Series Analysis tools to analyse gait data. In terms of medical signals, these tools were used for analysing heart and brain waves. However we employed these tools for video gait data and show the nonlinear nature of fronto-normal gait, which is an new and important observation. We then used a particular nonlinear approach, namely chaos theory in a novel application. This is the use of a measure of chaoticity, to characterize gait and use it as a biometric. Natural phenomena seems to exhibit this chaotic tendency, rather than the standard linear model in current gait analyses. This paves the way to explore the wide range of nonlinear analyses available. By combining chaotic characterization with face recognition, we demonstrate the efficacy of our approach to multimodal biometrics. Second, in the field of computer vision we provide a new comparison of methods for looming adjustment, as a subject approaches a camera. In handling objects which become self occluded, we employ signal processing tools in a novel application, that is, by casting it as a “missing data” problem. This allows for complete occlusion compensation over several frames of video data. This may be compared to current approaches to occlusion which ignore the problem, handle partial occlusion or only a few frames of full occlusion. Third, we gave a new and compelling case for fronto-normal gait as a biometric in its own right, by comparing our approach with that used in on currently available datasets. This was done with a qualitative analysis. Thus we show that fronto-normal gait has great potential to be used as biometric in its own right, and demonstrate the use of advanced time analysis tools to characterize it. By combining this with face recognition in an experiment on an early version of our dataset, we provide a platform for the deployment of biometrics in a very practical, useful and robust way.