Human visual perception inspired background subtraction

Background subtraction is an essential processing task for moving foreground detection. Existing approaches are reliable only when scenario-specific configuration is possible; otherwise, they exhibit highly unpredictable performance across a wide range of dynamic scenarios due to extensive dependence on statistical observations and context-specific constraints while lacking means for exploiting perceptual characteristics of the operating environment. This thesis investigates a theoretical framework for developing a novel human visual perception inspired background subtraction technique for unconstraint video analytics. The technique, named perceptual mixture-of-Gaussians (PMOG), emphasises on several perceptual characteristics of observed statistics for better exploitation of the operating environment to exhibit robustness in dynamic unconstrained scenarios. For instance, how human visual system perceives noticeable intensity deviation from the background; what is the perceptual tolerance level in distinguishing distorted intensity measures; and how realistic predictions can be made regarding an observation are the key questions investigated in the thesis. Addressing these questions has enabled PMOG to ensure high performance stability with superior detection quality across dynamic scenarios as well as optimal computational resource utilisation throughout the system lifetime. PMOG is then modified to improve responsiveness in unconstrained scenarios by incorporating a low-cost estimator of suspected foreground activities. The detection quality of PMOG is further enhanced by integrating two independent hypotheses originating from the same underlying model to maximise complementary aspects and minimise computational overhead. Comprehensive experimental evaluation is performed to establish superiority of PMOG against the state-of-the-art. Finally, the efficacy of PMOG is validated in the application domain of event detection.