Research On Visual Target Tracking Based On Kernel Correlation Filtering Algorithm

Visual system is the main way for humans to obtain external information.According to statistics,about 80% of external information is received through the human eyes.However,human vision is greatly restricted in various fields.With the rapid development of computer technology,allowing computers to process and improve human visual information has successfully promoted the generation and development of computer vision.As an important branch of computer vision,visual target tracking has a wide range of applications in the fields of intelligent video surveillance,robot navigation and positioning,human computer interaction,and virtual reality.In recent years,many target tracking algorithms have been proposed.However,target tracking still faces many challenges such as deformation,illumination change,scale change,occlusion and so on.When faced with challenges,one single algorithm may lead to the loss of target tracking.In response to these problems,this paper has made a more in-depth study based on the kernel correlation filtering target tracking algorithm.The main research content includes the following aspects:(1)For the research of kernel correlation filtering tracking algorithm theory,the correlation filter,ridge regression classifier,kernel function,cyclic shift matrix and cosine window function are studied.The cyclic matrix is used for ridge regression classifiers under dense sampling,and the fast training,fast detection and fast nuclear related processes can be obtained.The whole process of kernel correlation filtering tracking algorithm is summarized.(2)In order to solve the problem of kernel correlation filtering tracking algorithm poor tracking accuracy when tracking target faces deformation and illumination changes during the tracking process.The grayscale feature,CN(Color Name)feature and HOG(Histogram of Oriented Gradient)feature of the target are extracted,and then the weighted feature fusion method is used to integrated the three features together.The experimental results show that this method makes up for the shortcomings of poor tracking accuracy when the kernel correlation filter tracking algorithm uses only a single feature,and improves the algorithm's performance in dealing with target deformation and illumination changes.(3)In order to solve the problem of kernel correlation filtering algorithm drifts in tracking frame when tracking target scale changes during the tracking process.A scale adaptive strategy is proposed.The target position of the current frame is sampled,and different scale samples can be obtained.An independent scale filter is constructed,and the scale filter is learned and trained.The maximum value of the scale filter response value is selected as the optimal scale of current frame.The experimental results show that the scale adaptive strategy can solve the problem of target scale change in the tracking process.(4)Aiming at the problem of kernel correlation filtering algorithm model updating error when tracking target is occluded during the tracking process.A high confidence model updating strategy is proposed.The maximum value of response graph and the degree of oscillation of the response graph are calculated,and their historical average values are compared to decide whether to update the target model.The experimental results show that this high-confidence model updating strategy reduces the cumulative error generated when the model is updated in the occlusion background and improves the real-time tracking performance.