Head Tracking Based On Color Image And Depth Iamge

Target tracking in video sequences is a hot spot in the field of computer vision, and has broad application prospects in many fields, such as precision-guided, video monitoring, video compression, etc. In the current, safety monitoring is more and more attention,in schools, libraries, shopping malls and other public places, the monitoring system has become the essential facilities. Realizing intelligent monitoring, not only free human from repetitive tasks, but also can improve the security of monitoring.After years of research and development, there has been much more mature algorithm in the field of target tracking, but the occlusion problem has always been a difficulty in the field. Due to the loss of depth information, that only depending on the2D image cannot fundamentally solve the problem of occlusion in the tracking process. However, the traditional depth information retrieval algorithm and equipment is too complex to meet the needs of real-time tracking. In2010, Kinect depth camera launched by Microsoft provides a new opportunity for study in the field of tracking. In this article, considering the probability of head occuluded in dense crowd is less than human body, so this article will choose head as tracking target.This paper proposes a head tracking method which combines color image and depth image. The head tracking is modeled as an association problem of tag target and the candidate target. Our method extracts the foreground from the depth image firstly. Then the heads are located from RGB foreground image by a pre-trained head detector. The tracking is done by calculating the similarity between the detection responses and labeled targets. The feature of the heads is described as a combination of depth information, position information and the sparse-based appearance information. The depth information is the major feature as it can solve the ambiguity caused by2D plane image. The position and the appearance information obtained from the RGB image are employed as minor features to improve the reliability of the tracking. The proposed tracking method is able to handle many difficult problems such as crosses between targets, emerged-occlusion between targets and background, the reappearance of the disappear targets, and the severe occlusions in crowd scene. The tracking results on challenging test RGB-D image sequences demonstrate the effectiveness and the robustness of our method.