Research Of Moving Object Tracking Method Based On Kernel Correlation Filter

Moving object tracking is a popular topic in the field of computer vision,and it is used in intelligent transportation,military,unmanned and so on far and wide.There are endless researchers on target tracking algorithm at home and abroad,but due to the complexity of target tracking environment,such as motion blur,fast motion,and occlusion etc.,which can lead to significant changes in the appearance of the tracking target model,so to design a robust tracking algorithm is still a huge challenge.Gradually in recent years,the research on object tracking is moving from traditional algorithms to learning based algorithms,among them,the target tracking algorithm based on correlation filtering is regarded as the classification problem of target tracking,and the position of the target where confidence is the largest in the classification result is the position of the current target,it has a fast computing speed,and real-time learning and updating of classifier and target appearance.Therefore,this method has high tracking accuracy while keeping track speed,which has attracted wide attention of researchers.In this paper,the advantages and disadvantages of the object tracking algorithm based on correlation filter are discussed,and the algorithms are further studied.The main innovations are as follows.(1)When the target is moving fast,the target position of adjacent frames varies greatly,and the traditional kernel correlation filtering algorithm cannot solve such problems well.So on the basis of correlation filtering,this paper proposes a new detecting model,under the condition of ensuring algorithm time complexity,the search area is expanded and the target search is carried out in a more reliable area.In order to further improve the tracking performance of the algorithm,we propose a spatial constraint method,which is to assign a weight to each small search area.We assume that last frame tracking results have certain reliability,the next frame target tracking results may appear in more than one frame as the center of the search area of the probability is bigger,so it gives a higher weight to the small search area at the center.(2)In order to solve the problem of traditional kernel related filtering target tracking algorithm cannot handle well target occlusion,based on the new detection model,an adaptive template updating strategy based on occlusion detection is introduced.The new occlusion detection mechanism uses the historical frame information as the primary judgment of the occlusion of the target,secondly,the maximum response value and response mean of the current frame are used to determine the current occlusion of the target,then the target model is adaptive updated with the result of occlusion detection.When the target is not occluded,update the model according to the initial weight,and reduce the update weight of the model when the target is occluded,thus,the non-target occlusion information is few updated to the target model which makes the model update more accurate.Based on the various difficulties and challenges in the current target tracking,this paper makes an in-depth study on the target tracking algorithm of kernel related filtering,according to the existing defects,the corresponding improvement method is proposed.The data set used in the experiment is the standard video library in the field of tracking Visual Tracker Benchmark.The experimental shows that the proposed algorithm has better tracking performance under the fast motion and occlusion of target object.