Development of a Motion Distillation Paradigm for Visual Surveillance

Abstract

The huge number of cctv cameras and security applications places increasing requirements on automatic visual tracking and behaviour classification systems. The best working example of such a tracker is the human visual system (hvs) which can flawlessly detect, track and understand almost any object or event. The research described in this thesis uses lessons learnt from studies of the hvs to develop a novel approach for computerbased visual tracking. In this approach, initial detection of moving objects is achieved using a new motion distillation paradigm which employs spatio-temporal wavelet decomposition of video. The method is shown to be more robust than traditional background modelling techniques while being computationally less expensive. As with the hvs, the approach uses a dual-channel tracking architecture to perform tracking. The motion channel, generated through motion distillation, handles object detection and initialises tracking. The form channel is used to resolve tracking ambiguities and occlusions. Qualitative and quantitative tracking results illustrate the advantages of this approach. This thesis also describes a new approach to the task of ob- 4 ject (e.G. Human) behaviour analysis - a subject which is of great importance, yet which is still an under-researched aspect of visual tracking. In the work described here, objects are categorised into vehicles, pedestrians, runners, groups and unknown pedestrian behaviour.