Anti-occlusion Object Tracking Based On Feature-redetection

Visual target tracking is an important and difficult research point in computer vision.Due to the challenge of occlusion,target attitude scale change,camera jitters and fast target movement,there are still potential and space for many existing algorithm to improve.Among many tracking algorithms,TLD(Tracking-Learning-Detection)tracking algorithm uses the strategy of cooperation between tracking module and detection module to track,and introduces PN learning method to adjust on-line to achieve long-term tracking of single target,but this algorithm is prone to misjudgment when the target is occluded.This paper focuses on the analysis and research of TLD target tracking algorithm.The TLD algorithm is made up of tracking module,detection module and learning module.The tracking module predicts the position of the target in the next frame by means of median optical flow method;the detection module uses sliding window scanning and cascade classifier mechanism to screen out the possible position of the target in the next frame;the learning module uses P-N learning mechanism for online learning,and uses delicate integration strategy to synthesize the results of the two modules for decision-making,and obtains the tracking target in the current frame.The final position is updated and corrected by using the result.Based on TLD algorithm,this paper proposes a new method of anti-occlusion target tracking(TLD-D)based on feature re-detection.By introducing feature re-checking link,the problem of misjudgment caused by similar appearance of target and background clustering when occlusion occurs is solved.The strategy of tracking,detection,learning and re-detection is adopted.Combining tracking with detection,we can learn the latest appearance features of the locked target,and judge whether there is occlusion according to the calculation status of the current frame by tracking module and detection module.When occlusion occurs,we can extract two key frames of occlusion process: occlusion start and occlusion end.In the "occlusion start" frame,the preprocessing of feature reexamination is carried out to record the target state before occlusion start;in the "occlusion end" frame,the feature re-examination is carried out.The SIFT feature is used to match the positive and negative two-way feature to calibrate the target,and the pre-occlusion state is used to calibrate the target to ensure that the re-calibrated target is the original tracking target,and the results of feature re-examination are used to update the tracking and detection model.Block,namely "re-detection".Finally,the experiments on the target tracking benchmark OTB(Object Tracking Benchmark)show that compared with TLD algorithm and other classical tracking algorithms,the proposed TLDD algorithm has stronger anti-occlusion ability,better robustness and can track the target stably for a long time.