Research On Long Time Single-target Tracking Algorithm Based On TLD Framework

In today, effective use of computer tracking and precise positioning of the target is a hot topic in the field of computer vision. It has a wide application prospects.,For example used in military class navigation, traffic control, real-time video surveillance, diagnostic medical science, intelligent electronics, security and other areas of public life.TLD tracking algorithm framework of this study is a algorithm that based on a tracking, detection and learning. It is suitable for the unknown target tracking for a long time. In this algorithm, the detector and tracker are well together.it is in order to be able to keep on tracking results for evaluation and learning. The TLD algorithm has good robustness, and it also has good learning ability to recover. TLD tracking algorithm framework has good result without light interference characteristics and it also has certain anti shielding ability. But in the process of tracking if the target appears rotation or More than 50 percent of occlusion occurs, In that case,it will occur target is still in the camera but it failed to detect. The detector is the use of sliding window to detect classification, so need increase detector detection accuracy and reduce time complexity.Firstly, Improvement of the tracker that Based on the TLD framework. The improvement needs improved tracker can avoid the target rotation after tracking failure to some extent, In this paper, the tracker by using SIFT algorithm to ameliorate the issue of the target rotation after tracking failure. When the target can not be tracked, we need Matching with the feature of SIFT firstly. If the match is successful description of the target does not disappear just appeared rotation. Then you need to learn the new target in order to able to accurately track when it appear again. Otherwise we need dispose with the principle of the median flow.Second TLD detector algorithm was improved to reduce the time complexity of the detector.The method which plays an important role in detection accuracy and time complexity is how to reduce the sliding window which does not contain the target child into the cascade classifier TLD detector-how to send the child window which contains the scan target tracking into the cascade classifier efficiently.In the full study of the particle filter, the paper attempted to apply to theTLD detectors of a reasonable particle filter function and designed TLD detection algorithm based on particle filter.Finally, the paper describes the evaluation method of target tracking.The comparative analysis of the experiment proved that the algorithm can avoid the situation that target’s rotation leading tracking failure.The detector’s detection accuracy and time complexity are improved.