Research On Object Tracking Algorithm Based On Correlation Filter And Deep Learning

Single object tracking is a very important branch of computer vision.Single object tracking can be divided into short-term object tracking and long-term object tracking.Shortterm object tracking focuses on accurate object tracking in short sequence,while long-term object tracking emphasizes long-term stability,ability to deal with target disappearance,exit,loss and recovery on the basis of short-term task.The current single object tracking algorithm still has many shortcomings.For example,for the online update of tracking,online update is necessary because the target may undergo drastic appearance changes.However,it is difficult for the tracker to judge whether the tracking result of the current frame is correct or not.Offline training is of great help to the performance improvement of the tracker.Single off-line training is often inseparable from large-scale high-quality data sets,and the acquisition cost of such data sets is very expensive.In order to solve these problems,this paper propose the following three tasks:(1)For the background noise problem in the target box,this paper presents an adaptive spatial correlation regular correlation filter tracker model.The model can learn a correlation space regularization while learning the filter,which can effectively suppress the learning of background noise in the target box.At the same time,a positioning and scale estimation strategy is proposed to minimize the computation,which makes the algorithm achieve a real-time speed of 28 fps while achieving high performance.(2)Online update is a very challenging task for long-term tracking.This paper propose a long-term tracking algorithm based on meta-learning update controller.Firstly,an update controller is proposed,which encapsulates the discriminant information,geometric information and appearance information of the target in time sequence,and then sends it into the long and short time memory network to learn whether the current frame can be updated.Secondly,a long time tracking framework is proposed to make the short time tracker easily embedded into the long time tracker.The framework consists of the short time tracker,the full graph detector,and the validator.It can greatly improve the accuracy of updates,thus improving tracker performance.(3)In order to solve the problem of high cost of obtaining large-scale high-quality data,this paper designed an annotation algorithm to obtain high-quality complete annotation through sparse manual annotation.The framework proposes a selection and refinement strategy to automatically improve the preliminary annotations produced by the tracking algorithm.A TAssess Net is proposed to capture the temporal coherence of the target location and select reliable tracking results by measuring the quality of the target location.The VG-Refine Net is also designed to further enhance the selected tracing results by considering the target appearance and temporal geometry constraints,allowing inaccurate tracing results to be corrected.This algorithm can save about 94% of manpower when producing high quality annotations.