FAST TRACKING OF OBJECT CONTOUR BASED ON COLOR AND TEXTURE

Abstract

This paper presents an efficient algorithm to track the object contour in image sequences. Firstly, the probability distributions of color feature and texture feature are approximated by different schemes. Secondly, the object and the background pixels are distinguished by calculating their posterior probability in light of Bayesian formula based on the two models. Finally, an energy functional is proposed and solved by gradient descent flow to evolve the initial contour which is estimated by motion feature, so that minimum energy contour corresponds to the object contour. In order to save the computational resource, a new texture descriptor is applied to analyze pixels merely in the neighborhood of a contour rather than the whole image. Furthermore, the algorithm's implementation based on the narrow band level set scheme can help to reduce the processing time further. In some experiments, the contours of various objects such as human, car and body organ are successfully tracked by our algorithm and the processing time is short enough for practical application.