Research On Target Tracking Algorithm Based On Correlation Filter

Target tracking is one of the research hotspots in the field of computer vision.It is widely used in civil areas such as robotics,human-computer interaction and security,as well as military fields such as precise guidance of weapons.Tracking algorithm based on correlation filter has attracted the attention of researchers for its advantages of both speed and accuracy.In recent years,it has also achieved remarkable results.But at the same time,it still faces many challenges.In complex scenarios(such as occlusion,fast motion,etc.),the accuracy of tracker using low-dimensional feature is insufficient,and the algorithm using high-dimensional feature will reduce the real-time performance and affect the use effect.Based on target tracking framework of the correlation filter,this work studies boundary effect,scale estimation and model pollution in tracking.The main research results are as follows:Aiming at boundary effect,an adaptive extended search area method is proposed to alleviate boundary effect by shortening the distance between the search center and the real target center.Firstly,the least squares polynomial fitting is used to estimate the displacement according to historical location information of target and comfirm target motion state.When a large motion of the target is predicted,the search area is extended by introducing the second search area which may be closer to the real target.And then target detection is carried out in expanded search area.The experimental results on OTB datasets show that the proposed adaptive extended search area method can effectively alleviate the influence of boundary effect,improve the accuracy of target tracking,and maintain the processing frame frequency similar to the original algorithm in fast motion scenarios.Aiming at scale variation,this work proposes a scale optimization method based on Kalman filter,which is based on the correlation filter tracking algorithm of adaptive scale variation.The purpose is to improve the tracking performance of the algorithm by improving the accuracy of scale estimation.Firstly,the target scale transformation model is established.During target tracking,based on the optimal control theory,the Kalman filter is used to adjust the scale prediction of the scale transformation model by using the scale output of the scale filter,so as to complete filtering optimization of the scale measurement and obtain the optimal estimation of the target scale.The experimental results on OTB datasets show that the proposed scale optimization method can effectively improve the performance of the tracker while maintaining similar processing efficiency.Aiming at the problem of model pollution,a new adaptive model updating method is proposed in this work.Firstly,a method based on statistical probability is used to calculate the confidence score that the location area belongs to the real target,then the learning rate is adjusted according to the confidence score,and finally the model is updated.The experimental results on OTB datasets show that the proposed adaptive model updating method can effectively alleviate model pollution and greatly improve the performance of the tracker at the slight sacrifice of algorithm speed.In this thesis,boundary effect,scale estimation and model pollution are studied,and lightweight improvements are made to improve the performance of the tracker while maintaining a high processing frame rate,which provides effective support for the real-time application of the algorithm.