Target Tracking Algorithm Based On Discriminative Appearance Model

In recent years, as the application of video tracking technology is extremely broad, target tracking algorithm has attracted more and more attention from researchers. The key of the visual target tracking is to search for an appropriate representation method for object appearance, which not only can effectively extract appearance characteristics, but also can robustly estimate the motion state under the changes of background and target. This paper introduces the middle-level visual features as the basic unit of image description, and integrate the multi-scale superpixel into the target tracking framework based on discriminative appearance model. This method treat target tracking problem as a classification issue between target and background. The discriminative appearance model can effectively overcome the situation of similar characteristics between target and background, improving accuracy and robustness of tracking algorithm. Specifically, this paper studied the following three aspects:First, this paper analyzes the existing discriminative target tracking algorithm, in which the image processing unit can determine to a large extent of the performance. Although discriminative appearance model in a certain extent can distinguish the background and target effectively, the the feature expression based on pixel has a certain limitation. Therefore, this paper proposes a multi-scale superpixels discriminative tracking method from the perspective of mid-level vision with structural information captured in superpixels.Secondly, based on the above analysis, this paper proposes a multi-scale superpixel discriminative tracking method. We consider superpixel as the basic unit of image processing, training the Adaboost classifier, building the discriminative appearance model. Therefore, the tracking problem is transformed to distinguish two types of superpixels. And we compute a target-background confidence map based on the integration of multi-scale feature information while tracking, and obtaining the best candidate by scale meanshift. Tracking results is returned for online updating of discriminative appearance model.Finally, for a series challenging video sequences, we comparise our method with several existing methods. Experimental results demonstrate that our tracker is able to handle occlusion, large shape deformation, etc.