Physics-based evolutionary strategies and dynamic sensor fusion for moving object detection

Moving object detection remains a fundamental problem in computer vision. Difficulties arise due to the changing environment, the environment in which a moving object operates, where the background is highly dynamic. Slow illumination changes due to sun position in the sky or sudden variations due to clouds passage for example, create challenges for all moving object detection systems including feature-based and featureless methods. Another fundamental challenge in all moving object detection systems is the problem of distinguishing between the moving cast shadow from the moving object itself. For robust detection, in addition to identification and removal of the moving cast shadows, the dynamics of the scene must be modeled and tracked (adapted) appropriately. This dissertation addresses these two fundamental problems through introduction of a novel technique that overcomes limitations in the existing moving object detection techniques. The proposed methodology has the following innovative features: (1) It provides several competitive and cooperative fusion strategies for moving object detection. (2) It introduces a new physics-based shadow detection algorithm based on sound physical models that can incorporate a variety of background and foreground surface types. (3) It includes performance analysis of the physics-based shadow detection algorithm on a variety of objects and backgrounds. (4) It integrates thermal and reflectance physical models into a uniform approach for sensor fusion. (5) It develops a novel sensor fusion technique based on cooperative coevolutionary computational model, which incorporates physical and environmental conditions into a new evolutionary dynamic fusion model. (6) It provides analysis and results of moving object detection using color and IR video for a full diurnal cycle.; The efficacy of the methods presented is demonstrated through extensive experimental evaluations on real data of both visible and thermal images for various surfaces and environmental conditions.