Research On Moving Target Tracking Method Based On Kernel Correlation Filter In Occlusion Environment

Object tracking is a basic research direction in the field of computer vision.Profit from the rapid development of computer graphics processor and memory,the technology has been applied in many fields in real life,such as human-computer interaction,intelligent transportation,security video monitoring,aerospace and sports analysis.The correlation filtering target tracking method regards the tracking problem as a classification problem,and determines the target position by taking the maximum confidence point of the classification results.In recent years,the correlation filtering target tracking method has been widely concerned by scholars in the field of computer vision due to its excellent real-time performance and tracking accuracy.Due to the complexity and diversity of the actual tracking environment,the moving target may deform,change in aspect ratio,blur or exceed the field of vision.In the process of camera shooting,there are also situations where the target moves too fast or the target is occluded,which brings great challenges to the correct tracking of the target and greatly limits the accuracy of the target tracking method.Therefore,in this paper,the correlation filtering method is improved for target deformation,fast movement and occlusion problems,and the improved algorithm is verified by experiments.The main contents of the paper are as follows:A kernel correlation fast target tracking algorithm based on position one-step prediction scale adaptive is proposed to solve the problems of deformation,blur and lost tracking caused by fast moving maneuvering targets in video sequences.Firstly,the gray scale and color feature of the current frame are extracted,and the gradient and color histogram of the gray scale feature and color feature are calculated to obtain the direction gradient histogram feature and color histogram.Then,according to the color histogram,particle filter is introduced to predict the target position in the next image,and then identify the search area centered on the predicted target location.Finally,the current target location estimate is modified based on the directional gradient histogram feature and kernel correlation filtering algorithm in the search area.On this basis,the deformation and ambiguity problems are solved by combining the zero intercept and the average peak related energy.Through quantitative and qualitative experimental analysis,the algorithm can effectively track the fast moving target.Aiming at kernel correlation filtering method lacks the ability to deal with scale and occlusion,an anti-occlusion target tracking algorithm based on depth feature and scale estimation is proposed.Firstly,the depth features of Conv3-4,Conv4-4 and Conv5-4 layers of the current frame were extracted by VGG-19 convolutional neural network,and a position filter was trained in each layer.Then,the position filter was used to predict the position information of the target in different convolutional layers in the next frame.At the same time,corresponding weights were assigned to features of different convolutional layers according to the richness of semantic information,and target positions were obtained according to the location information and corresponding weights of targets in each convolutional layer,so as to update the position filter template.Then,one-dimensional scale filter is used to predict target scale based on target position.Finally,a residual function is designed to judge whether the target has occlusion at different positions of convolution layers.When the target is occlusion,the template update is stopped,and the SVM classifier is used for target re-matching in the following frames.Through quantitative and qualitative experimental analysis,the algorithm can effectively track the target with scale variation under occlusion.