Research On Improved Kernel Correlation Filter Tracking Algorithm Based On Multi-feature Fusion

At present,the rapid development of the field of computer vision has provided people with a lot of convenience.The human visual system has some limitations in acquiring and understanding information,and the computer can be used to capture and perceive information from the outside world.However,as the scale of video image data continues to expand,how to simultaneously meet real-time and accuracy is still a problem that many scholars need to solve.The target tracking field is a comprehensive technology of multiple technologies,which has the characteristics of wide application fields and large development space.At present,there is no general algorithm that can accurately track in any scene,so the target tracking algorithm has a lot of room for development.In recent years,related filtering algorithms have been proposed,which has greatly improved the real-time and accuracy of target tracking technology,and more and more scholars have developed a keen interest in this type of algorithm.Nevertheless,algorithms based on correlation filtering still cannot track error-free in all complex scenes.For example,when the scene has similar target interference,the tracker relies too much on the maximum response value,which leads to template drift and tracking errors.Considering the above problems,this paper proposes an algorithm for multi-feature fusion and improved kernel correlation filtering.First of all,in view of the problem that the training samples generated by the cyclic matrix are easily distorted,we have introduced a regularization matrix into the improved tracker to make the samples more robust in the training process.Second,we selected two complementary features to train the improved tracker.In particular,we introduced a re-detection module to provide more target candidates and relocate uncertain targets.We use the response threshold to start the detector module.When the response value is greater than the set value,there is no need to start the re-detection module;when the response value is less than the set value,the re-detection module is started to avoid tracking errors caused by the core-related filtering algorithm overly relying on the maximum response value.Finally,we took into account the more common scale change problems and fast motion problems,and adopted adaptive scale and template update solutions.We conducted a large number of experiments on the algorithm proposed in this paper in the OTB data set,using video sequences in six different scenarios for experiments,and compared with some representative algorithms.The first set of experiments compares the center position errors of each algorithm in different scenarios.The second set of experiments compares the running time and accuracy of the algorithms.In terms of accuracy and real-time performance,we calculated the average tracking of different algorithms.Accuracy and average number of frames per second.The higher the average accuracy and average number of frames per second,the more frames that meet the requirements,and the stronger the accuracy and real-time performance of the algorithm.Finally,it can be concluded from the effect graph of the experimental operation that the algorithm proposed in this paper has more advantages in tracking effects in complex scenes such as similar target interference.