Research On Target Tracking Algorithm On Kernel Correlation Filter

In recent years,target tracking in dynamic scenes has become an important research direction in computer vision,and has important significance in many applications such as robot technology and intelligent monitoring.Due to the change of position,posture and scale during the movement of the target,the occurrence of occlusion and the change of the background,there are still great challenges to accurately track the target,so how to improve the accuracy of target tracking in dynamic scenes has theoretical and practical significance.The classic kernel correlation filter(KCF)algorithm has been successfully applied to target tracking,but the KCF target tracking algorithm has the following problems:First,the size of the target frame remains unchanged during the tracking process,resulting in the background information being introduced or part of the target information being lost when the scale of the target changes;The second is the lack of adaptive threshold adjustment mechanism in KCF tracking,which makes it impossible to determine the loss of the target.The third is that the artificial features used by KCF lack the ability to describe the target in complex scenes.Therefore,we propose a series of solutions to the problems in the classic KCF target tracking algorithm.The main work of the improved kernel correlation filtering tracking method proposed in this paper includes:(1)An online updated nearest neighbor classifier is used to confirm the tracking target,so as to determine whether the target is lost,so as to improve the stability of tracking,and reduce the tracking complexity of the tracker and detector combination;(2)In order to solve the problem that the traditional KCF algorithm cannot obtain the constantly changing size of the moving target,the scale change of the target is obtained by calculating the ratio of the bandwidth of the candidate region centered on the spectral response peak in the front and back frames,In order to achieve the adaptive change of the size of the target during the tracking process;(3)The HOG feature of the original KCF is replaced with the depth feature extracted by the Resnet50 pre-training model based on target detection to form a tracking target template to achieve a stronger description feature of the target.In order to verify the effectiveness of the improved KCF target tracking algorithm proposed in this paper,the target tracking data set OTB's DarkCar,Basketball,Woman,Skater2 four data sets and the Mytestl,Mytest2 data sets we photographed were tested extensively.Experimental results show that,compared with the classic KCF tracking method,our method's success rate and accuracy on six data sets are improved by an average of 25.5%and 34.6%.In order to further verify the contribution of each module to the overall tracking performance,we also conducted ablation experiments.Compared with the classic KCF target tracking algorithm,the realization of adaptive change of target size,the use of nearest neighbor classifiers and the adoption of depth features have improved the average success rate by 2.34%,21.55%,and 4.05%,and the average accuracy has been improved by 5.20%,25.50%,5.72%.The experimental results show that the tracking method based on kernel correlation proposed in this paper can greatly improve the accuracy and success rate of tracking,so that it can be applied to the automatic tracking of targets such as video surveillance and robots.