Moving Target Tracking Based On Correlation Filtering In Complex Scenes

In actual production and life,target tracking technology plays an important role in our lives.It has very important application value and development prospects.At present,the target tracking method based on correlation filtering is still a hot research content in the field of target tracking due to its ultra-fast tracking speed and excellent tracking performance.In this paper,the research work is carried out based on the kernel-related filtering target tracking algorithm.The algorithm uses only one HOG feature in the tracking process,the tracking target frame remains unchanged during the tracking process,and the tracking target is easily lost when the tracking target is occluded.Improvements have been made,and the main research contents of this article are as follows:(1)Aiming at the problem that the kernel-related filter tracking algorithm uses only one HOG feature to track the target during the tracking process,which leads to unsatisfactory tracking results in many complex tracking scenes,an improved multi-feature fusion target tracking algorithm is proposed.According to the analysis of the characteristics of various common features in the image,the four complementary features of HOG feature,CN color feature,gray feature and LBP texture feature are finally selected to be fused.Considering the influence of multi-feature fusion on the tracking speed of the algorithm,we choose to combine the low-dimensional HOG features,CN color features,and gray features first to the serial fusion of feature layers to obtain new HCG features.Then,the HCG feature obtained by the fusion is fused with the LBP texture feature of higher dimension for the feature fusion of the response layer,so that only two correlation filters need to be trained separately to use all the four complementary features for target tracking.(2)Aiming at the problem that the target tracking frame always keeps the same size during the tracking process of the original KCF algorithm,it is easy to cause tracking failure in scenes where the target scale changes greatly.A correlation filter is proposed to track the target scale.After the location of the target is tracked through the feature fusion response,the scale pyramid technique is used to obtain n samples of different scales to train the scale filter,and the trained scale filter is used to correlate the HOG features extracted from the next frame of scale samples by calculation,the optimal scale of the instant target of the scale corresponding to the position with the largest response is obtained.(3)Aiming at the problem that the tracking accuracy of the kernel-related filter tracking algorithm is not high when the target is occluded during the tracking process,a kernel-related filter tracking algorithm fused with Kalman filter is proposed.First,judge whether the target is occluded during the tracking process.In this chapter,the peak side-lobe ratio PSR is selected as the method to determine whether the target is occluded.When it is decided that target is not occluded,the conventional correlation filter is selected to predict the target’s position.When it is judged that the target is occluded,the Kalman filter to assist tracking is selected to predict the target’s position,and the conventional filter stops updating at this time.The effectiveness of each chapter’s improved algorithm is verified experimentally on the OTB2013 target tracking data set.The experimental results show that compared with the original KCF algorithm,the improved algorithm has a certain improvement in tracking accuracy and tracking success rate,which shows that this chapter The improved algorithm is effective and feasible.