The Research On Tracking Algorithms Of Video Moving Objects

Moving objects tracking in video sequences are the fundamental and challenging research topics in computer vision area at present. They have great research and application values in modern industry, military, avigation and spaceflight area.While to now robust object tracking is still an open problem especially when the background is complex. The main obstacles in the research are the scale of object, occlusion, background complexity and non-rigid deformation. In this research our goal is find a real-time and robust object tracking algorithm by employing the novel object feature representation, layer idea, multi-scale analysis, filter and prediction.An effective moving object tracking algorithm is proposed by combining Kalman filtering method and template feature points matching. Firstly, Scale-Invariant Feature Transform feature points are extracted to represent the moving object, which are then matched with these feature points on the object template to obtain motion vectors of the objects. Secondly the motion vectors of the object will be processed by a clustering algorithm to reduce noise and mis-matching. Finally, the tracking algorithm is developed by Kalman filtering method on feature matching result. Experiments show that the SIFT features are invariant both to scale and rotation. It is insensitive to object deformation and illumination variation. The feature point clustering algorithm together with Kalman filtering is robust to scale variances and partially object occlusion. Despite of the precision and robustness of the above mentioned algorithm, its efficiency is very low since the feature extraction of SIFT has a very large computational cost. That is reason why we proposed the following modification.To obtain precise tracking result and tracking efficiency, an improved moving object tracking algorithm is presented using the multi-scale Histograms of Oriented Grads & Colors (HoGC) based on the framework of traditional Kalman Filter and Mean-Shift optimization. Firstly, based on the fusion of oriented grads and colors histogram a new HoGC is proposed, which is invariant to moving object's rotation and deformation. Secondly, HoGC pyramid is constructed to characterize the multi-scale objects more robustly. Then, by coupling kernel-based Mean-Shift algorithm with Kalman Filter, HoGC matching is optimized in the scale and position spaces and the position of the candidate object is identified. Finally, experiments on representative video sequences are accomplished. Experiments validate that the proposed multi-scale HoGC is robust to represent the object and is invariant to scale and deformation, and the proposed tracking algorithm not only can improve the reliability and accuracy which traditional histogram-based tracking algorithm cannot, but also can improve the real-time performance which SIFT tracking algorithm fails. The improved tracking algorithm is better than the traditional algorithm in accuracy and speed.Our work can service the applications of video analysis, intelligent surveillance, activity analysis and synthetic.