Research On Long-term Target Tracking Algorithm Based On Dual Detectors

As one of the research hotspots in the field of computer vision,target tracking has a wide range of applications in the current society.Although many excellent algorithms have been proposed,the performance of target tracking algorithms in some aspects is still not ideal due to the constraints of tracking challenges and training samples.Based on the related object tracking algorithm,this thesis has done the following work for the problems existing in the tracking process:In order to solve the problem that the target tracking algorithm cannot identify the target after the long-term occlusion,a long-term target tracking algorithm based on the dual-detector system based on the related filtering algorithm has been proposed.In the tracking process,when the reliability of the target position is low,the detector will be activated to re-detect the current area of picture.The detector is a combination of a support vector machine detector and a siamese network detector.The support vector machine detector is used to quickly detect at the current area.If no reliable target is detected,the siamese network detector is started for deep search.Finally,the target re-detection is achieved through the cooperative work of the two detector.In order to deal with the phenomenon of target drift during long-term tracking,an adaptive spatial regularization term is added,which can effectively suppress background information and increase the robustness of the model.At the same time,the alternate direction multiplier method is used to optimize the objective function to reduce the cost of time.Experiments show that the algorithm achieves the expected results on the OTB dataset,especially when tracking long video sequences,the improvement is obvious.In order to solve the problem that the target may re-enter from any other place after moving out of view,a target tracking algorithm based on the out-of-view detector is proposed.In the tracking process,if the target moves out of view,the detector will be activated;the detector monitors the edge of the picture in real time,and uses the maximum response value and the cosine correlation of the singular value matrix to determine whether there is a target in the current monitoring result.Then the detector is shut down until a more credible target is found.In order to train a more reliable filter,an improved average peak correlation energy value is proposed as a criterion for the current tracking results.If the tracking effect is good,the model is updated.Finally,by training two correlation filters to estimate the target scale and position respectively,the performance of this algorithm is getting better.Experiments show that the algorithm can achieve a relatively ideal tracking effect,and effectively deal with the situation where the target is out of view.Also it has a better performance in the face of other challenging factors.