Athletes In The Sports Video Detection And Tracking Method

Detection and tracking of athletes in sports video is very important because it can provide high-level sports video processing such as motion analysis, event detection, 3D reconstruction with necessary information. Based on the analysis of the current moving object detection and tracking algorithms, this paper researched the detection and tracking method of athletes in sports video as well as improved some traditional algorithms. In the end, a prototype system was developed to validate the effectiveness of improved algorithms.In order to extract moving regions in sports videos, a mixture of Gaussian model which was updated by an on-line approximation was adopted to model the video images, and then adaptive background subtraction was used to improve the accuracy of moving regions. The shortcoming of the traditional Adaptive Gaussian Mixture Model (GMM) is that it suffers from slow learning at the beginning, especially in busy environments. In order to deal with this deficiency, a novel update mechanism was introduced. Firstly, expected sufficient statistics update equations were used, and then the L-recent window update equations followed. In order to increase the accuracy of detected regions of athletes, a texture similarity measure was applied to detect and eliminate shadow, which was followed by extracting regions of athletes according to the information of detected playfield.The CamShift (Continuously Adaptive Mean Shift) algorithm was used to track athletes, which utilized color features to locate and track targets. The advantage of traditional CamShift is its robustness and good speed. However, because it only considered a single color channel (Hue) while modeling the athletes, it would fail to track the target when the Hue alone cannot allow the object to be distinguished from the background and other objects which have similar color distribution. In order to improve the robustness of CamShift algorithm, we presented a variational CamShift. We modeled the athletes by three color channels (Hue, Saturation, Value) in HSV color space, and then replaced the traditional histogram by weighted histogram and ratio histogram. Occlusion between athletes often exist in sports videos, so an occlusion detection scheme was also introduced in this paper. We compared the search windows at a fixed time interval, the rapid change of window size means existence of occlusion. When the serious occlusion was detected, the moving region detection algorithm was re-introduced to detect the athletes, which was followed by the athlete tracking algorithm.This paper was supported by the National High-Tech Research and Development Program of China (863 Program) (No.2006AA01Z328) and the Open Foundation of State Key Laboratory of Computer Science, The Chinese Academy of Sciences (No. SYSKF0704).