Application of Physics-Based Image Formation Models to Change Detection in The Context of Indoor Workplace Video Surveillance

The aim of this thesis is to investigate the application of physics-based image formation models to change detection in the context of indoor workplace video surveillance.

First, video surveillance applications are reviewed. Based on this review, a new classification of video surveillance applications is proposed and indoor workplace surveillance is chosen as the target application. A new workplace surveillance modelis then introduced, which relates the needs of workplace surveillance applications, their implications and the capabilities of video surveillance techniques. Furthermore, a set of requirements for workplace surveillance applications are elicited and a videobased workplace system structure is proposed.

Change detection is then reviewed, and the suitability of using physics-based image formation models to enhance change detection algorithms is investigated. Two physics-based change detection techniques are developed. The foundations of these techniques are advances of colour constancy techniques which extract physical features from the camera output; this approach is unlike other change detection algorithms which use the camera output directly without considering its physical meaning.

The performance of the proposed techniques is measured and compared against the Horprasert algorithm, using objective and computational complexity evaluation methods, where the quality of the change detection is measured using recall and precision measures. The Horprasert algorithm was shown, in an independent study by other researchers, to have the best trade-off between segmentation quality and computational complexity among other state-of-the-art algorithms such as Cavallaro, McKenna and Shen under experimental conditions which covered different lightings and background structures.