Research On Motion Target Tracking Algorithm Based On Kernel Correlation Filter

Computer vision is a science that allows computers to learn to "see" instead of the human eye,and is also the study of how to make artificial systems "perceive" from images or multi-dimensional data.As an important task,target tracking has important application value.There are many factors that affect the stable tracking of the target,and it is a major challenge to stabilize the tracking under any influence.The correlation filter tracking algorithm has attracted the attention of many researchers with its high tracking accuracy and excellent processing speed.However,many excellent tracking algorithms have unsuitable scenarios and factors that can affect their tracking effects.For example,the tracking effect is not ideal when the target has a similar background or is blocked.Aiming at these two problems,this paper proposes a tracking algorithm fused with SIFT feature matching based on kernel correlation filtering.First of all,considering that the kernel correlation filtering algorithm with a single feature cannot adapt to complex backgrounds or different backgrounds,this paper introduces the sift feature.When the original single feature is not enough to accurately describe the tracking target,the sift feature is extracted to more accurately represent the target feature;Secondly,the kernel-related filtering algorithm lacks occlusion processing.After the target is occluded,the tracking template is easily contaminated in the model update process,so that the target frame drifts and the target is lost in the following tracking process.In this regard,the original update mechanism was changed in the experiment.When the target is blocked,the template update is stopped until the target reappears and the template parameters continue to be updated,reducing drift in the tracking process and making the tracking process more stable.In order to capture the target more accurately when the target is reproduced after the target is occluded,this paper adopts forward and reverse feature matching to obtain the optimal feature point after the match.Mark in the first frame after the occlusion,eliminate the sparse points and keep the dense points,circle the target at the dense feature points and continue tracking.This algorithm can find the target again when the target is occluded,and solves the problem of no recheck after the target is lost in the kernel-related filtering algorithm.It can also accurately detect the target when the target is confused with the background.In the experiment,first extract the HOG feature of the region of interest and find the largest target response based on the kernel correlation filter for tracking;when the target is occluded or the tracking frame drifts due to similar background influences,the SIFT features of the target area are extracted for feature matching and tracking,and stop updating the template;use SIFT feature matching to get the re-checked target position,and then continue stable tracking.The experimental results show that the algorithm can solve the problem of target frame drift caused by the target being temporarily occluded and the target and background confusion to a certain extent.