Object tracking via inverse sparse representation and convolutional networks

Abstract

In this paper, we present a novel object tracking method based on inverse sparse representation and convolutional networks. First, in contrast to existing trackers based on conventional sparse representation, the target template can be sparsely represented by candidate dictionary in our method and the candidates corresponding to nonzero coefficients are selected as the some optimal candidates of tracking results. At the same time, locally normalized features are adopted to obtain the representation of target template and candidate dictionary, which can deal with partial occlusion and slight object appearance change. Second, a convolutional network is proposed to select the best candidate from the candidate set got by inverse sparse representation. Numerous bank filters are introduced to preserve local structure of the target and background samples and the feature maps are extracted to form the simple layers and complex layers. Finally, a simple local model update scheme is employed to accommodate occlusion and target appearance change. Both qualitative and quantitative evaluations on several challenging video sequences demonstrate that the proposed method can achieve favorable and stable results compared to the state-of-the-art trackers.