Research On Scale-adaptive Target Tracking Algorithm With Anti-occlusion

Target tracking is an important research in the field of computer vision,and has a wide range of applications in the fields of intelligent robotics,intelligent transportation and video surveillance.The correlated filtered target tracking algorithm has become a milestone in tracking algorithms in recent years because of its excellent accuracy and speed advantages.In practical scenarios,the change of target scale and the occlusion of the target by other objects can have a large impact on the accuracy of target tracking.In this paper,we propose a scale-adaptive anti-occlusion target tracking algorithm based on the classical Kernelized Correlation Filters(KCF)target tracking algorithm to address the problem of low tracking accuracy due to target scale change and target occlusion during the tracking process.can effectively improve the tracking accuracy when the target scale changes and the target is occluded.The main research work and results are as follows.(1)A scale-adaptive target tracking algorithm based on depth estimation and feature fusion is proposed to address the problem of low tracking accuracy due to the change of target scale in complex backgrounds.The algorithm uses a depth estimation network to get the depth of the target,and establishes a depth-scale estimation model by combining the scale value of the target;in the tracking process,the depth estimation network is used to get the depth value of the target,and then the scale value of the target is obtained according to the depth-scale estimation model,so as to achieve scale-adaptive target tracking.In addition,Histogram of Oriented Gradient(HOG)and Color Name(CN)features are used to jointly build the appearance model of the target and improve the robustness of the model.Experiments using this paper's algorithm and several classical algorithms on five video sequences with target scale changes in the public dataset OTB-100 and one self-constructed video sequence show that the experimental results are similar to those of KCF algorithm,KCF algorithm using HOG features and CN features(Kernelized Correlation Filters with CN,KCF-CN)algorithm and Discriminative Scale Space Tracking(DSST)algorithm,the average accuracy indexes are improved by 21.86%,18.59%,and 6.21%,respectively,and the average accuracy indexes are improved by 21.86%,18.59%,and 6.21%,respectively,compared with those of KCF algorithm,Kernelized Correlation Filters with CN(KCF-CN)algorithm,Discriminative Scale Space Tracking(DSST)algorithm,and Spatially Regularized Discriminative Correlation Filters(SRDCF)algorithm,Integrate 1D Boundary and 2D Center CFs(IBCCF)algorithm,the speedup is 0.904f/s,8.836f/s.(2)A depth estimation-based anti-obstruction target tracking algorithm is proposed to address the problem of low tracking accuracy due to target occlusion in complex backgrounds.The algorithm uses correlation filtering templates to perform correlation calculations with candidate targets,and judges whether the current tracking is reliable based on the obtained maximum response value.The unreliable candidate targets are chunked and the occlusion detection is performed using the maximum difference of the depth values of each chunk.When the target is partially occluded,the range of the candidate area is reduced,and when the target is completely occluded,the range of the candidate area is expanded.And when the target is occluded,the filter template is not updated,thus effectively preventing the filter template from being contaminated and avoiding the subsequent tracking from drifting.The experiments are conducted on four video sequences in the public dataset OTB-100 where the target is occluded using this algorithm and the comparison algorithm.tracking(LCT)algorithm,the average improvement in accuracy metrics is 28.85%,19.925%,and 18.15%,respectively.