Research On Object Tracking Based On Position-Scale-Feature Space Coordinated Correlation Filtering

Object tracking based video is widely used in intelligent transportation,security monitoring,human-computer interaction and other fields.However,the tracking effect is often affected by scale change,obstacle occlusion,target deformation,moving out of view,background clutter and other factors.Therefore,how to overcome these difficulties and achieve fast and robust target tracking is still worthy of further study.Based on the correlation filtering object tracking algorithm,this paper proposes the following three improved algorithms from the aspects of feature extraction,sample fusion,model update and scale estimation as follows:Firstly,aiming at the problem that the traditional correlation filtering tracking algorithm is easy to fail when the target changes scale and occlusion,this paper proposed a multi-feature selection correlation filtering target tracking algorithm based on position-scale spatial coordination.The features of fast histogram of oriented gradient,color names and gray of the region are extracted then different features are fused to form feature pools for correlation filtering tracking to enhance the discriminant capability of tracker.The response map with the highest robustness score of feature response is selected to predict the target location.The phase correlation filter is used to estimate the target scale in log-polar coordinates,updating filter template with the strategy of high confidence model updating based on the tracking effect.Secondly,aiming at solving the problem of the boundary effect caused by fast motion and the template error learning under the circumstances of occlusion,this paper proposed a robust tracking algorithm based on multi-feature joint spatial-temporal regularization.The algorithm extracts the hand-crafted feature and the deep convolution feature of the target region,and dimensions of the features are reduced by using principal component analysis.The temporal domain and spatial domain regularization terms are added to the correlation filter tracking framework to optimize the solution,and the feature scale pool is used to estimate the scale of the tracking target adaptively.Finally,aiming at the traditional correlation filter tracking framework is limited by the single resolution feature and the scale estimation is not precise enough,a multi-scale estimation and adaptive response fusion target tracking algorithm based on convolution feature is proposed.In this algorithm,multi-resolution convolution features are fused by learning convolution operators in continuous domain,and Gaussian label parameter is adjusted to give full play to the robustness of deep features and the accuracy of hand-crafted features.The algorithm speed is improved by decomposing convolution operation,dynamic sample fusion and fuzzy sparse model updating,and the object response map is adaptively fused according to the index of prediction quality evaluation.