The Research On Object Tracking Algorithm Based On Compressive Sensing

Target tracking technology has become one of the main research and application aspects with the rapid development of related technology in the field of computer vision,but tracking arbitrary target stably and accurately in complex scene, there still exist many theoretical and technical problems to be solved, so developing a robust online tracking algorithm is still a very challenging task.The tracking algorithm which based on the principle of compressed sensing not only alleviate the tracking drift problem to some extent, and it can reach very good real-time effect that doing better than other online tracking algorithms.My paper research is based on compressive tracking algorithm, the main work is as follows: how to enhance the apparent model robustness and select the better classification features, as well as developmenting a search strategy that need less calculation.1.Making a analysis of current compressive tracking algorithm, and understanding its mechanism,making a comparison between it and other online tracking algorithms through testing to proved its stable tracking performance and good real-time effect.In the end, to summarizing the advantages and disadvantages of this compressive tracking algorithm, pointing out the next step work target.2. In a complex scene, due to the intense light, severe occlusion and so on, object appearance may changes much in a new frame. In order to making the appearance model to deal with these changes better, this paper puts forward a improved compressive tracking algorithm which combining kalman prediction and sub region adaptive updating together. The sub region classification parameter is allowed to update when its output value changing extent meet the update threshold condition. At the same time, the algorithm using the tracking result from the minimum output change of the sub regions to figure out the whole target location, but when the minimum classification change value exceeds the threshold used for disturbance judgment, then using the kalman prediction value as the target output for current frame.3.Making an improved feature selection method and an improved sample search strategy together, the former is based on feature distribution distance which can distinguish target,and the latter method making the amount of calculation reduced.Both the two methods can improve the accuracy and the robustness,the improved search strategy significantly reduces the computation that can make up the increased computation by offline feature selection. On the one hand, the selection of optimal feature, the feature distribution of the original compressive tracking satisfies Gauss distribution,then select the method based on the distance of feature characteristics probability distribution of positive and negative samples. On the other hand, using a simple and efficient tracker search two positions in every frame by a coarse to fine search strategy.4.Do experiment on the kalman sub-region feature update method, and the improved feature selection and search strategies method, comparing the effect of tracking algorithm in this thesis to the original algorithm, experiment results demonstate that the improved tracking algorithm has better robustness.