Research Of Target Tracking Algorithm Based On Vision

Along with the development of scientific technology, image processing technology rapid development in recent years, computer vision has become very popular in the field of the computer. And target tracking as it essential and important part, has been used widely to such as human-computer interaction, intelligent traffic, robot vision navigation and military, etc. For instance, in intelligent transportation system, through to tracking vehicles for traffic statistics for the intersection with a busy traffic flow can let the traffic police department work reasonable arrangement, avoid traffic jams; In the safety monitoring system, through to tracking the personnel to on their behavior analysis, judge whether for suspicious people; In car auxiliary driving system, through to tracking the vehicles on the road, get the vehicles’ information such as position, velocity, backing to the system to calculate the out ahead vehicle’s speed and distance, timely report to drivers avoid a collision. How to track target accurately and rapidly in every complicated scene is becoming a significant research subject. Therefore, the main contributions in this thesis are as follows:(1) Firstly this thesis has researched and analyzed the Lucas-Kanade optical flow method, the Mean Shift algorithm and trust region method three popular traditional tracking methods. They have a common characteristic is that just considering the tracking target and not considering the information of around the target, this thesis called it as a absolute tracking method. Based on the general use of the color matched the method, which makes them unable to solve such as complex scene, illumination changing and occlusion etc. difficult problem in the process of tracking. In addition, optical flow method is based on the hypothesis of two frames in the same brightness, so can’t handle the case such as target change; Mean Shift although processing speed faster but can’t solve scale change; And trust region method for tracking effect is good, but it is difficult to reach a real-time effects.(2) Compared with the traditional tracking algorithm, this thesis adopts a novel based on relatively tracking algorithm, which takes into account both to the target itself and around the target background information. Online random forest algorithm is one of the relative tracking method. Online learning and detection-based method of tracking is the most main features. According to less of positive sample in online learning, this thesis proposes an improved algorithm of online random forest. In addition, in choosing on the training features, this thesis uses a compatible with simple and illumination unchanged features-BRIEF, and according to the characteristics of target tracking this thesis makes the corresponding improvement.(3) In view of traditional detection method is based on the map exhaustively method, needs to spend a lot of time, it is difficult to meet the multi-objective real-time tracking. This thesis proposes to combine the Kalman filter with random forest, estimating of possible target position through the Kalman filter, only detecting in the near scope. This thesis improves the detection method, reduces a lot of unnecessary times, making the system more easily do real-time target tracking.At last, this thesis realizes the tracking algorithm, and contrasting with traditional methods, the experimental results show that this method is better than the traditional tracking methods.