Research On Video Target Tracking Method Based On Deep Learning

With the advancement of technology and the explosive growth of video surveillance data,video target tracking has became a hot area of computer vision.It is widely used in video surveillance,navigation detection,human-computer interaction and military fields.However,many difficulties have been encountered in the actual tracking process,such as obstructions,changes in lighting,complex backgrounds and analog interference.Among the many challenges,the most troublesome one is occlusion.For example,the popular kernelized correlation filter tracking algorithm in the past two years is to directly apply the occluded samples to the classifier when the occlusion occurs in the target,resulting in the weakening of the classifier performance.Then tracking drift occurs.In order to improve the tracking performance of the kernelized correlation filter tracking algorithm when the target is blocked,this paper improves on this basis and proposes a video target tracking method based on deep learning.In order to solve the problem of loss caused by occlusion,a target occlusion detection module is established firstly.According to the idea that target forward tracking and back tracking should be consistent,a target occlusion detection algorithm based on one-step backtracking is proposed.This algorithm determines the number of frames back in one step by the rate of overlap between historical tracking frame and the current frame and then determines whether the target has occluded and the degree of occlusion that is a partial block or a serious block by obtaining the rate of the backtracking box and the target real frame overlap.Considering that the occlusion first takes into account the reference of historical training samples,this paper proposes a target-tracking algorithm for kernel-related filters based on historical training samples.This algorithm combines the number of historical samples with target samples to expand the number of samples and reduce the effect of negative to classifier performance.However,its effect is better for local occlusion,but its corresponding performance is still not good when the target is seriously occluded and lost.In order to better performance optimization,this paper is subject to deep learning in the wide range of applications in the field of target detection in recent years and inspired by the unique advantages of deep learning networks.ResNet,a deep residual network,is applied to the tracking algorithm for target loss detection.In addition,the tracing data characteristics of traditional ResNet is improved to realize robust long video target tracking.Finally,in order to verify the feasibility and robustness of the video object tracking based on deep learning proposed in this paper,the method is compared with TLD,Struck,STC and KCF which are popular in video target tracking methods.Firstly,the setting of occlusion threshold,the number of historical training samples and the output dimension k of ResNet network are set up through experiments.Then,under the condition that the experimental environment,experimental data set and experimental evaluation index are the same,through a large number of experimental results contrast,quantitative analysis and qualitative analysis prove the validity and robustness of the video target tracking method based on deep learning proposed in this paper.