Long-term Moving Target Tracking Method Fused With Multiple Features

In recent years,the subject of moving target tracking has occupied a place in the direction of computer vision.However,general algorithms cannot cope well in some complex scenarios such as target deformations,rotations or occlusions,the target might get drift or lost and r esulting in failure of tracking,so,accurate and real-time tracking of targets has always been a problem that needs to be solved in th e field of computer vision when targets are interfered by various factors.In this paper,some mainstream tracking algorithms have been studied and analy zed,and some improved methods are proposed to solve the problems often encountered in moving target tracking,which achieves better results and has been verified.This paper aims at color video sequences,the conventional Kernel-Correlation Filter tracking method does not take the color information into consideration and lack of occlusion processing scheme when the target is under the influence of occlusion,changes of lighting,etc.,there might be a large tracking error and a low tracking precision.In order to solve this problem,following points are proposed: Firstly,combine the HSV spatial color feature with the HOG feature,to deal with some problems such as deformation or rotation of the target.The HSV color feature describes the global feature,while the HOG feature describes the local feature of the target,the combination of the two features can give a more complete description of the target;Secondly,an adaptive scale estimation method is used to estimate the scale of the target,which reduces the interference information that might be caused by the scale transformation of the target,so as to a chieve the purpose of accurate recognition and description of the target and improve the robustness of the algorithm;Finally,to deal with the problem of losing of the target,an online random fern classifier is used for redetection,and the patch with th e highest confidence is regarded as the new location of the target.In the experiment,some challenging video sequences in the public test video sets are selected,and the results show that the algorithm proposed has a strong robustness in some complex scenarios such as scale transformations,morphological changes,occlusions,rotations and so on.