Single/multi-target Tracking Algorithm Based On Cell Correlation Matching

With the rapid development of computer vision,target tracking technology has been widely used in civil and military fields,and has become a hot research field.In tracking tasks,there are many challenges,such as target occlusion,temporary departure field of view,illumination change,target deformation,etc.,which often cause tracking task failure.Therefore,it is of great research value to improve the performance of the tracker in dealing with complex situations,optimize tracking strategies,and improve the accuracy of single target or multi-target tracking.In this thesis,the main research work is divided into two parts(1)A target tracking algorithm SiamMeta based on region similarity matching is proposed.Aiming at the problem of target attitude change in long-time tracking task,the Siamese network algorithm is upgraded from single template to multi template to enrich target features;Aiming at the challenges of target occlusion and temporary departure from the field of view,the cross-correlation calculation method is changed,and the whole feature matching is changed to "meta" feature matching to find the uncovered part;Aiming at the problem of "meta" matching frame drifting,a new method is adopted.The anchor frame regression method can more accurately return to the target frame;For the problems of fast movement and motion blur,the optical flow assistant system is introduced to distinguish the target and background through the optical flow diagram,to correct the prediction frame.After the verification of OTB2013,OTB2015 and other data sets,the algorithm surpasses the original baseline in many indicators(2)Based on the framework of Deepsort multitarget tracking algorithm,KCF-deepsort and SM-deepsort multitarget tracking algorithms are proposed.KCF-deepsort upgrades the prediction module from the classical Kalman filter algorithm to the KCF tracking algorithm of real-time update tracker,and improves the prediction accuracy with the help of KCF;SM-deepsort uses the SiamMeta based on the deep learning template that is proposed in this thesis as the prediction module,combined with the twin network optimization data association module.This paper analyzes whether the template should be updated in the multitarget tracking task,and experiments the two algorithms on Cityflow cross camera public dataset and airborne dataset respectively.The results show that SM-deepsort,which uses SiamMeta as the prediction module without updating the template,performs better in dealing with complex situations.