Research On Object Tracking Based On Spatio-Temporal Information

Object tracking is one of the most important problems in the field of computer vision with applications ranging from surveillance to human-computer interactions. Numerous factors affect the performance of a tracking algorithm, including illumination, variation, occlusion, and background clutters. Although object tracking has been made in recent years, it is a challenging problem to making a strong tracker of robust performance, fast processing speed. To improve the robustness of the object tracker, a novel tracking method which track target via spatio-temporal information was proposed.In this paper, the research work carried out for two aspects: single-object tracking and multi-object tracking for static single camera:(1) To better solve the occlusion issue, a novel tracking method which track target via part spatio-temporal information was proposed. The task of tracking was decomposed into translation and scale estimation. For the translation estimation, the target and its components were become tracking objects. Firstly, the feature that can represent object nicely was chose adaptively. Then the statistical correlation between the feature from the object and its surrounding regions was modeled. By learning the correlation, the respective spatio-temporal information was got, which can be used to track the target and its components. Finally, the target’s direction, speed, color information were modeled by a sliding window, to evaluate the performance of each object tracker. Based on the evaluation results, the most reliable tracking object was chose. The final center position of target was estimated with the target architecture model and the position of the most reliable tracking object. For the scale estimation, a multi-scale target pyramid at the center position of the target was built to estimated the optimal scale of the target. Extensive experimental results show that the proposed algorithm favorably against state-of-the-art methods in terms of efficiency, accuracy, and robustness.(2) In this paper, a multi-target tracking algorithm based on spatio-temporal information and tracklet confidence was proposed, which can effectively deal with issues such as missing detection. The task of multi-object tracking was divided into local and global association.In the local association, tracklets with high confidence were locallly associated with the provided detections, whereas fragmented tracklets having low confidence were globally associated with other tracklets and detections. The matching score was determined by the affinity model. The spatio-temporal information was added to the affinity model, which make it become more robust and overcome the missing detection problem in tracking-by-detection methods. Moreover, the criteria of Smooth Constraint of Confidence Maps and Peak-to-Sidelobe Ratio were used to measure the reliability of temporal information. Experiments with challenging public datasets show distinct performance improvement over other tracking mathods.(3) The purpose of multi-object tracking algorithm for video sequence is to calculate the position of all the objects in each frame image. In this paper, a framework to achieve the task of multi-object tracking by a single-object tracking was proposed. The framework extended the single-object tracking algorithm based on part temporal-spatial information into a multi-target tracking. The main difficulty lies in: Firstly, how to determine the object is a new object; Secondly, how to judge an object have left the observation window; Thirdly, only single-object tracking may not track objects correctly, so when the trajectory of the whole video sequence was got, analyzing and updating the trajectory must be done. Inspired by linear discriminant analysis, the difference between tracklets and the similarity in tracklet were used to define a tracklet association cost, whose variance was used to find and correct wrong trajectory. Experimental results show that the proposed method of tracklet adaptation can effectively detect and correct wrong trajectory.