Visual tracking for arbitrary shaped object

The accuracy of observation model to represent of natural object is critical to performing robust visual tracking in complex scenarios. In this paper, we propose an object observation model by considering three integrated feature components, i.e. the template component depicting the object structure in every time instant; the contour component characterizing the object shape evolution with local deformation, and the colour distribution component accounting for long-term unstructured appearance. The corresponding tracking algorithm is comprised of a rigid patch matching in inter-frames and a region-based level set segmentation for local contour transformations. Experimental results on real complex situation demonstrate the performance of the proposed tracking method.