Correlation Filter Based Real-Time Object Tracking

In recent years,correlation filter based algorithms have been successfully applied to the object tracking problem and have shown excellent performance.Traditional tracking algorithms run fast but suffer from poor accuracy and can only be applied to simple scenarios.The mathematical models of many new tracking algorithms are very complex and even require the integration of deep learning for visual feature extraction,which causes they suffer from heavy computation cost.Based on the correlation filter algorithm,this paper proposes an efficient object tracking algorithm to track multiple arbitrary objects in real time on the CPU side.The tracking algorithm improves the accuracy and robustness through spatial constraints and feature weights.The motion estimation model is introduced to alleviate the influence of occlusion and severe deformation.The final filter response is analyzed to evaluate the tracking quality for dynamic update of the underlying models through multiple learning rate fusion method.In addition,the proposed algorithm fully uses pre-computed Fourier transformation,dynamic memory management,parallel computation and feature dimension reduction to do speed up.Furthermore,an asynchronous delay update strategy was proposed to reduce the computation cost.Experiments on the VOT dataset have shown excellent testing results.Compared to current object tracking methods,the proposed algorithm achieves 6%improvement in tracking results and 20%improvement in tracking speed.The frame processing time can be reduced to,milliseconds while maintaining good tracking quality,which brings wider application space.