Research On Correlation Filter Tracking Algorithm Via Sparse Update Strategy

Object tracking has always been one of the hot topics in the field of computer vision.Its purpose is to predict the subsequent motion trajectory of a target based on the known target position at the current time.In the actual scene,because of the changes of the target itself,background information interference and occlusion,the tracker will introduce background noise into the tracker template due to inappropriate model update strategy,which will greatly weaken the tracker's recognition ability and The anti-interference ability to noise leads to template drift and the final tracking fails.In this paper,the target tracking algorithm is studied in terms of confidence discrimination,model update strategy and tracking re-detection of tracking results in occlusion environment.The main work of the thesis is as follows:(1)The model drift caused by the frame-updated template update strategy adopted by the Staple tracking algorithm greatly affects the robustness of the target tracking algorithm.Based on a large number of observation experiments,we found that the sparse model update strategy can effectively avoid the model drift and improve the speed and robustness of the tracker.This paper proposes a sparse update strategy based on the maximum response value in the period.Firstly,the confidence of the tracking result in the current period is judged based on the variation law of the maximum response value of the tracking result in a certain period;secondly,the target model is updated based on the tracking result with the highest confidence in the period.The experimental results show that the proposed algorithm outperforms the performance of the mainstream target tracking algorithm,especially when dealing with occlusion and severe deformation.The algorithm based on the OTB10O benchmark dataset has an AUC of 0.633,which is a 9%improvement over the underlying algorithm Staple.(2)When the target is occluded,since the background information greatly interferes with the apparent characteristics of the target,the target information cannot be stably tracked only by the feature information of the target itself.Therefore,based on the related filtering target tracking algorithm,an SVM detector is integrated in parallel,and the''simple"samples and"complex"samples are labeled based on the self-step learning mechanism,and the positive and negative samples in the historical video frame sequence are used to continuously update and optimize.SVM detector.When the confidence level discriminates that the video tracking result is unreliable in the current period,the SVM detector re-detects all the video frame sequences in the current period,compares the re-detection result with the tracking result of the correlation filter,and compares the confidence.High results are recorded as the actual location of the target.Finally,the optimal result is selected to update the detection classifier and the correlation filtering model respectively.Experimental results show that the algorithm further improves the performance of the target tracking algorithm in dealing with occlusion problems.The algorithm has an AUC of 0.643 in the OTB100 benchmark data set.This paper compares the current mainstream tracking algorithms and compares them with the OTB 100 benchmark dataset.The experimental results show that the proposed two algorithms have better tracking performance than other comparison algorithms.