Research On The Deformable Target Tracking Algorithm

Target tracking is an important research direction in computer vision.Target tracking has many practical uses in video processing and is a classic computer vision problem.In real life,target tracking is used in video surveillance,human-computer interaction interfaces,robot perception,motion recognition,medical imaging,and drones.In the process of target tracking,when the target is deformed,the traditional target tracking algorithm can no longer meet the tracking requirements.Therefore,this paper studies the problem of tracking accuracy degradation or tracking failure caused by target deformation.The main research contents are as follows:Aiming at the situation that the apparent model trained by a single feature cannot adapt to the change of target appearance resulting in tracking failure,this paper proposes a multi-feature fusion target tracking algorithm.First,calculate the difference between the HOG and CN features in response to the PSR of two adjacent frames to obtain the fusion weight of the two features.Use the obtained weights to adaptively fuse the response of the HOG and CN features.The response and the response obtained by the color histogram feature are subjected to secondary fusion with fixed weights.Finally,the target center position is determined by calculating the maximum value of the final response.Through testing on the dataset OTB50,the results show that the apparent model established by the algorithm in this paper can adapt well to the change of target appearance,and the average accuracy of the algorithm is improved by 2.2% compared with the previous algorithm.Aiming at the problem that the filter template brings more error accumulation due to target deformation and the tracking accuracy decreases,this paper proposes an improved filter adaptive update algorithm.First,classify the target tracking credibility by combining the range of difference between the PSR of the target response value of each frame and its mean,thereby judging the tracking effect,and secondly,learning the position-dependent filter and scale-dependent filter based on the determined tracking effect The rate is dynamically adjusted.Finally,compared with nine algorithms such as DCF?CA and CSRDCF,the results show that the algorithm in this paper is superior to other algorithms in multiple performance indicators,and reduces the error accumulation of the template.Aiming at the problem that the traditional target frame can not accurately represent the size of the deformed target,which will cause the target tracking success rate to decrease,this paper proposes a target rotation adaptive tracking algorithm.First,the target is converted from a rectangular coordinate system to a polar coordinate system,the Fourier Merlin features of the target are extracted in the Fourier domain,and then a rotation sample set is established according to the extracted features,a rotation filter is trained,and the target rotation angle is estimated.Finally,the experimental verification shows that the rotation filter constructed in this paper can calculate the target rotation angle well,more accurately represent the target size,and the accuracy and success rate of the algorithm have also been improved.