Improved Target Tracking Algorithm Based On TLD

Computer vision as the current hot field,its core problem of video target tracking is a key and difficult point of scientific research.The video target tracking algorithms existing have various kinds and advantages,but in practice,it is generally faced with problems such as tracking target occlusion,tracking target appearance changes and environment mutation.This paper focuses on analyzing and studying the tracking algorithm of Tracking-Learning-Detection(TLD).The TLD combines the thought of target detection and target tracking,and divides the whole system into three parts: tracking module,detection module and learning module.The tracking module uses Lucas-Kanade(LK)optical flow method to predict the position of the tracking target in the next frame.The detection module performs a global scanning and cascade classification for the next frame to determine where the tracking target may appear in the next frame.The learning module uses P-N learning mechanism to learn the result of tracking module and detection module to update the cascade classifier in the detection module,so as to improve the robustness and stability of the TLD.In addition,the comprehensive part synthesizes the result of tracking module and detection module to determine the final position of target in the next frame.Based on the TLD,this paper improves the TLD to solve the problems on the occlusion and scale change of the tracking target.For tracking module,this paper proposes a method of combining corner detection and uniformly taking to choose the tracking points,giving full consideration to the integrity of tracking target and the disturbance of background information.Tracking target is divided into the main part of the basic part with no background information and edge part with background information.The basic part adopts the theme of the uniformly taking to get the number of stable tracking points and the edge part uses corner detection to get some reliable tracking points.The method of combing corner detection and uniformly taking ensures that the number of tracking points are selected to avoid the subsequent LK optical flow method to track when no points available,and to some extent,improves the quality of the tracking points to improve the probability of success tracking by subsequent LK optical flow method.Further,this method improves the robustness of tracking module in TLD.In order to make up for the amount of calculation in the choosing tracking points method of combing corner detection and uniformly taking,this paper sets Markov predictor in the bottom of tracking module,considering the result of Markov prediction and normalized cross-correlation predictor decides scanning area of detection module that greatly reduces the amount of calculation compared to the original global scanning.For the detection module,this paper puts forward a prior scanning method with similar scales.Usually,the intense deformation in tracking target between two adjacent video images won't appear,so firstly using the three scanning windows that are more similar to the size of target bounding box in the current frame to scan and do cascade classification in the next frame.The detection module according to the result of cascade classification decides the location of the object in the next frame.Compared with the original scanning window according to the growing up scales of scanning strategy,a prior scanning method with similar scales effectively reduces the computation of detection module,further improving the real-time performance of TLD.