Research And Implementation On Kernelized Correlation Filter Tracking Algorithm

With the development of computer vision technology,visual object tracking as one of the basic topics in computer vision is taken more and more attention by researchers.Visual object tracking has a wide range of applications in human-computer interaction,video surveillance,mobile robot and so on.In recent years,the discriminant tracking algorithm that transforms the tracking problem into a detection task becomes a hot spot of research since its good performance.Subsequently,researchers applied correlation filter in classical signal processing to discriminant tracking algorithm,and reduced the complexity of training and detection.However,a single object tracking algorithm couldn't cope with all situations,especially in complex visual scenes such as occlusion,fast movement,fast jitter of the lens,and so on.In thesis,we focuses on the object tracking algorithm based on kernelized correlation filter,and proposed a tracking algorithm based on complementary fusion.Then we implemented it on software.In thesis,the main research and contribution are as follows:(1)In thesis,we first study the discriminant tracking algorithm based on regularized least-squares classifier,including the training and detection process in linear case and kernelized case.Then we applied the dense sampling method to regularized least-squares classifier,and reduced the complexity including fast training,fast detection and the fast kernel correlation.Finally,we obtained the process of kernelized correlation filter tracking algorithm.(2)Since the limitation of kernelized correlation filtering algorithm in the complex visual scene,we combined kernelized correlation filtering algorithm,spatio-temporal context tracking algorithm and the online detection algorithm into a tracking framework,and proposed a complementary fusion tracking algorithm.Then we simulated it on 18 video sequences of complex scenes,and the average distance accuracy is 94.61%,which is 26.27% higher than kernelized correlation filter algorithm and 43.54% higher than spatio-temporal context tracking algorithm.The average success rate is 83.32%,which is 24.63% higher than kernelized correlation filter algorithm and 43.66% higher than spatio-temporal context tracking algorithm.(3)In thesis,we implemented the complementary fusion tracking algorithm on PC platform with C++ language,and used the performance analysis tool of visual studio-profiler to analyze the performance of the program.Then,to improve the efficiency of program,we used the skills of the program optimization to optimized the hot part of the program.The experimental results show that the average frame rate is 9.2 FPS(frame per second)before optimization and increased to 31.5 FPS after optimization.In the real scene,the experiment shows that the optimized program has a good tracking performance on the object with serious occlusion.