Compressed Sensing Theory And Its Application On Visual Tracking

As one of the fundamental subjects in computer vision, visual tracking plays a critical role in numerous lines of advanced technologies such as object identification and behavior analysis. It has been widely applied in the fields of mobile robots and intelligent video surveillance. It is still a challenging job to develop a robust online tracker due to difficulties to account for the appearance change of a target, which include intrinsic (e.g., pose variation and shape deformation) and extrinsic factors (e.g., varying illumination, camera motion and occlusions). The tracking algorithm based on compressed sensing theory can deal with heavy occlusions, appearance deformation, varying illumination and motion blur very well. The researches on compressed sensing theory and its application on visual tracking have been studied in detail. The main work of this paper is as follows:(1) The compressed sensing theory is studied in detail. Simulation experiments on natural image reconstruction have been done to verify the compressed sensing algorithms and the framework. Also, experiments have been done to compare four representative reconstruction algorithms (BP, LASSO, OMP, St-OMP) in the aspects of reconstruction accuracy and the time complexity. St-OMP is proved to consume the least time on the same conditions and therefore has value to be further studied.(2) The object tracking algorithm based on compressed sensing theory is thoroughly studied. Experiments have been designed to compare the tracking accuracy of L1 Tracker, Incremental Learning Method (IVT), Positive and Negative Samples Learning (P-N), and Robust Frags-based Tracking (Frag-tracker). L1 Tracker is verified to have higher tracking accuracy and stability between these algorithms.(3) A combination algorithm of IVT algorithm and L1 Tracker framework is studied. Firstly, RPCA theory is introduced into the L1 tracker’s framework, and PC A eigenbases are treated as the target templates. The noise factor is put forward to estimate the noise degree. The dynamic template updating strategy based on noise detection is studied. Instead of the LASSO algorithm, St-OMP is applied to the framework in order to improve efficiency of the algorithm. Simulation experiments prove that the combination algorithm has higher tracking accuracy than L1 Tracker and IVT in various noisy circumstances.