Fast hybrid block/region-based algorithm for object tracking

This paper presents a technique to segment and track a region of interest in a video sequence. The presented technique starts with a spatio-temporal contour extraction. This contour extraction is performed by mixing information from a spatial gradient and a temporal differentiator. A block segmentation is then applied on the image, and within each block the contour part is approximated by a straight line segment. These contour approximations are then further quantized by a vector quantization technique, using a predefined codebook. In order to close and smooth the region of interest contour, an interpolation based on hermite polynomials is performed on the set of discrete quantized contour parts. Finally, the tracking process is performed using an addition-removal of blocks containing contour parts by testing their possible belonging to the polynomial contour. This testing is done by using a blurred version of the polynomial contour in order to be able to track deformations in the object. The initial guess for tracking is given manually by an end-user.