Reach On Object Tracking Algorithms Based On Spatial-Temporal Adaptive Kernel Correlation Filter

Thanks to the "Made in China 2025","13th Five-Year Plan" and other policies put forward,and the country's support for high-tech industries,target tracking has seen rapid growth in energy,5G,big data,AI and more.Relying on the needs of enterprises and society,computer vision is increasingly important in the field of intelligent security,personnel monitoring,vehicle detection,behavior recognition and other fields,target tracking as a key technology of machine vision,this paper,under the premise of improving the real-time target tracking algorithm,while improving the algorithm for target masking,deformation of the robustness of some work done.In recent years,discriminative correlation filter(DCF)has obtained extraordinary improvements in object tracking.However,the DCF method has a poor tracking performance on rapid deformation and fast motion.Spatially Regularized DCF(SRDCF)solves this problem by introducing spatial regularization to the filter at the cost of increasing complexity.The Background Aware DCF(BACF)used negative samples generated by the displacement of the real samples to improve the ratio of real samples.STRCF added temporal regular terms to the DCF.Besides,the tracking model of STRCF is more robust than SRDCF in the case of large appearance variations.In this work,we propose an adaptive spatial-temporal regularized correlation filter(ASTRCF)with a model update suppression strategy to balance temporal coefficients and spatial regularization weights,which can effectively learn the new appearance of the target and background over time and space.Firstly,we can get a closed-form solution by the alternating direction method of multipliers(ADMM),and some popular methods such as KCF,SRDCF,STRCF,BACF,and ASRCF can be regarded as an approximate solution in the particular scene of the proposed method.Secondly,we employ a model update suppression strategy to calibrate the misjudgments in the process of updating the model and to avoid model damage.Finally,we perform comprehensive experiments on four benchmarks:VOT2017,UAV123,OTB-2013,and OTB-2015.Moreover,our tracker runs at over 25FPS.Numerical results show that our method has an improvement in accuracy and robustness compared with the state-of-the-art trackers.