Superpixel-Based Compressive Tracking

Object tracking under complex scenario is one of the hottest topics in the field of computer vision.After decades of research,object tracking technology has made great progress and has a wide range of applications in the civilian and military fields such as the video surveillance,intelligent transportation,human-computer interaction.However,object tracking remains a challenging problem in handling complex object appearance changes caused by illumination,pose,occlusion and cluttered background.These factors put forward high requirements on the robustness and real-time performance of the object tracking algorithm.Currently,tracking methods based on the compressive sensing theory by applying random measure matrix can extract the low-dimensional features and greatly enhance the tracking speed and are attracting more and more attentions.However the tracking results may become inaccurate if the samples of the background and the foreground target have similarities in the shape or texture and accordingly the strategy may not work.For this,this paper proposes the superpixel-based compressive tracking(SCT)algorithm,which constructs a confidence map for a new frame according to the similarity of the superpixels between the new frame and the first frame in the surrounding region of the target.SCT incorporates the merits of superpixel segmentation that groups pixels into perceptually meaningful atomic regions.The confidence map provides strong evidence to measure the possibility of the target appearing,which captures the differences of the local appearance between the target and the background at superpixel level and improves the coarse-to-fine search strategy of FCT algorithm.In conclusion,this paper proposes the superpixel-based compressive tracking algorithm.Not only do SCT consider the different shape or texture between the target and the background,but also they take full advantage of discriminative color descriptors as a guidance.Experimental results on challenging sequences show the proposed algorithm outperforms state-of-the-art algorithms in terms of accuracy and robustness.