Human tracking method based on multi-template color-texture mean-shift algorithm

Human tracking is a hot topic and a challenging task during the past few decades. This paper present a multi templates based strategy for human detecting and tracking with a mobile robot. This method first determines the coarse location by using adaptive template matching algorithm (ATM) based on head-shoulder. Then, a multi-templates based method is presented to locate the person precisely. Multi templates considering the pose changes are obtained to represent the person. For each template, the mean-shift is proceeded. Then, the accurate position is obtained by fusing the results of the Mean-shift from all the templates. After detecting the person, the templates are updated by considering the likelihood of the tracking results and the old templates. Finally, the method is evaluated on a mobile robot in complex environment. The experiment result shows that our method performs well when there are unclear disparity image and pose variations.