Research On Object Tracking Based On Sparse Representation

Object tracking has been a significant research direction in the field of computer vision. It is carried out by analyzing object features and matching the most similar features in every frame. However, developing a robust tracker is not so easy, maybe confronted with many challenges and unknown problems, such as illumination change, scalar change, occlusion and target pose change and so on. Sparse representation is a popular research in recent years, and this paper mainly studies two robust tracking algorithms based on sparse representation.Firstly, combining object global features and local features into generative tracking framework, this paper proposes a tracking method based on a collaborative model. This method uses the advantage of both global appearance model and local appearance model, and, according to the different states of the image sequences, chooses suited tracking model in a dynamic way. This method uses incremental SVD to update global model, and according to projection error of object and feature space, uses a dynamic update strategy in updating local model. In this way, our method avoids the mistakes caused by occlusion in model update effectively.This paper then proposes a tracking method based on multiple instances learning in discriminative tracking framework. Considered the advantage of local sparse representation in describing the object features, and the good performance of MIL classifier in classification, local sparse codes are used to describe object features and MIL classifier is used to train sample data. Under the particle filter framework, this method first uses the dynamic classifier to get the initial object and then uses static classifier to locate the object state in current frame.By doing a lot of experiments, the tracking algorithms perform well against other state-of-the-art methods, and both qualitative and quantitative evaluations demonstrate our methods have a good robustness and stability.