Particle filtering on large dimensional state spaces and applications in computer vision

摘要: Tracking of spatio-temporal events is a fundamental problem in computer vision and signal processing general. For example, keeping track motion activities from video sequences for abnormality detection or spotting neuronal activity patterns inside the brain fMRI data. To that end, our research has two main aspects with equal emphasis - first, development efficient Bayesian filtering frameworks solving real-world tracking problems second, understanding temporal evolution dynamics physical systems/phenomenon build statistical models them. These facilitate prior information to trackers as well lead intelligent image understanding. The first part dissertation deals key challenges involved. In simple terms, basically estimating hidden state system noisy observed data(from sensors). As frequently encountered real-life, due non-linear non-Gaussian nature spaces involved, Particle Filters (PF) give an approximate inference under such setup. However, quite often we are faced large dimensional together multimodal observation likelihood occlusion clutter. This makes existing particle filters very inefficient practical purposes. order tackle these issues, have developed implemented on applications various visual vision. In second dissertation, develop dynamical inspired by human cognitive ability characterizing pattern shapes. We take landmark shape based approach representation activities. Basically, change configuration “landmark” points (key interest) over time use automatic extraction tracking, sequences. this regard, demonstrate superior performance Non-Stationary Shape Activity(NSSA) model comparison other works. Also, owing space problem, utilized filters(PF) tracking. third algorithm able presence illumination variations scene. do learn 2D Legendre basis functions. Under formulation, pose task joint motion-illumination thus PF called PF-MT(PF Mode Tracker) addition, also change/abnormality framework across drastic changes. Experiments real-life usefulness while many approaches fail. The last explores upcoming field compressive sensing looks into possibilities leveraging ideas better sequential reconstruction (i.e. tracking) sparse signals small number random linear measurements. Our preliminary results show several promising it interesting direction future potential applications.