Multi-cue Fragment-based Human Tracking For Mobile Robot

Human detecting and tracking for a mobile robot equipped with a vision system is an important topic in computer vision. And it is widely used in many areas, such as surveillance and motion capture, detection and tracing intruders and so on. However, there still remains a challenge, due to the problems caused by occlusion, illumination change and distinguish people. To solve these problems, a human tracking system based on the multi-clue fragment-based human tracking for mobile robot was proposed. Furthermore, an associating visual graphical user interface was developed. The method was verified by experiments conducted on a Pioneer mobile robot equip with a Kinect system. The research work mainly included the following aspects:(1) The coarse location with compressive featureGenerally, the tracking algorithms are discriminative models, which can separate the target from background by a classifier. But in the complicated environment, the algorithms are difficult to satisfy the requirements of real-time tracking due to the large calculation. To deal with the problem, a compressive sensing based algorithm is presented to get a rough estimation of the object location. Meanwhile, learning parameter of the compressive classifier is adjusted adaptively to avoid over-updating based on the maximum similarity of compressive feature and the average similarity of color and depth features.(2) Multi-clue fragment-based human tracking algorithmThe traditional tracking algorithm with color- texture feature cannot effectively handle illumination change, occlusion and so on. In order to resolve the issues, an efficient and robust tracking method by using multi-clue block matching is presented in this paper. In the proposed algorithm, compressed feature obtained by sparse measurement matrix is adopted to obtain coarse position of the target. Based on the results, the block matching is adopted to establish the precise position, combining with the color, depth and motion clues. Furthermore, depth clue can eliminate the affection of the illumination change.(3) Updating Sub-block Weight and Target ModelThe traditional tracking algorithm cannot effectively handle invariable target model and occlusion. In the proposed algorithm, the target color and depth model are updated by considering the feature contribution which determine by the similarity of color and depth features. The update scheme can not only eliminate the accumulated error, but also avoid drifting away. And, the traditional block matching algorithm doesn’t consider the differences between the respective sub-blocks which often results in tracking failure. In this proposed algorithm, the sub-block weight is adjusted adaptively according to the similarity of color and depth features which can ensure the robust tracking.In conclusion, the human detection and tracking algorithm described above is verified on the Pioneer3-DX robot platform equipment with a Kinect. Experimental results show that the presented algorithm, which incorporate multi-clue in the block matching, can satisfy real-time requirement and enhance robustness of robot tracking.