Research On Target Tracking Algorithm Based On Compressed Sensing

Object tracking has been extensively researched because of its importance in practical applications including medical imaging, traffic monitoring, visual surveillance, etc. Despite extensive study on this topic with demonstrated success, numerous issues remain to be addressed. It is still a challenging task to develop a robust and efficient tracking system to account for various appearance changes caused by occlusion, scale changes and illumination changes. In this thesis, the key technology of object tracking is deeply studied, and the theory of compressed sensing is introduced comprehensively. Object tracking, however, faces great challenges in practical applications, the robust object tracking algorithms are presented to address these challenges.An improved compressive tracking algorithm with scale adaption is proposed. The compressed features of objects in special scale are efficiently extracted by an adaptive random measurement matrix in compressed sensing domain. The tracking task is formulated as a binary classification via a naive Bayes classifier with online dynamic update in the compressed domain. The classifier is used to separate the target object from the background with particle filter framework. Furthermore, in order to improve the performance of our algorithm, a method is adopted to estimate the location of target by a combination of rough and accurate estimation.An advanced compressive tracking algorithm based on a weighted classifier boosted by feature selection is presented. The compressed features with high discrimination are selected from the target information of previous and current frames by a discrimination evaluation strategy. These discriminating features are used to train a weighted classifier, which is composed of two sub-classifiers based on previous and current samples bags. Finally, the weighted classifier is used to tell the target object from the background.Experimental results from various benchmark challenging sequences show that the proposed method runs in real-time and performs favorably in terms of efficiency, stability and robustness.