Research On Long-term Target Tracking Algorithm Based On Deep Learning

Target tracking plays an important role in intelligent surveillance,autopilot and military guidance.In recent years,with the continuous improvement of the performance of the deep learning short-term tracking algorithm,people began to pay attention to the long-term tracking applications close to the actual scene.The video sequence length of long-term tracking is much longer than that of short-term tracking,and the problems of target deformation,disappearance and reappearance are particularly prominent.The direct application of short-term tracking algorithm can not deal with these difficulties,and the tracking performance drops sharply.Therefore,this paper proposes the following two improved algorithms for deep learning longterm target tracking.Aiming at the problem of target disappearance and reappearance in the process of longterm tracking,this paper designs a siamese network long-term target tracking algorithm based on dynamic template matching(SiamDTM_LT).The confidence score is used to judge the target tracking status.If the target is lost,the global search and re-detection mechanism of dynamic template matching is started to obtain the rough positioning of the target.Then the SiamFC++tracker is used to accurately locate the target position,so as to solve the problem of target disappearance.In order to improve the accuracy of rough positioning during re-detection,an adaptive dynamic matching template updating strategy is also proposed.Tested in five longterm datasets of VOT2018_LT,VOT2019_LT,UAV20L,TLP and LaSOT,the results show that SiamDTM_LT algorithm not only has significantly improved the tracking performance,with a success rate of 0.556 on the lasot dataset,but also has a tracking speed of 45.5fps,which meets the needs of real-time target tracking.In order to further improve the tracking performance of the algorithm,a double model Siamese networks long-term tracking based on dynamic template matching(DMSiamDTM_LT)is designed in this paper.Two improved strategies are mainly proposed:(1)the "local-globallocal" tracking strategy improves the tracking stability of the algorithm when the target leaves the field of view and is partially occluded;(2)SiamFC++double model tracking strategy fully adapts to the changes of target appearance and improves the anti-interference ability of the algorithm.Tested in five long-term datasets of VOT2018_LT,VOT2019_LT,UAV20L,TLP and LaSOT,the results show that the tracking performance of DMSiamDTM_LT algorithm is significantly improved,and the success rate on LaSOT data set is 0.574.Compared with other advanced target tracking algorithms,DMSiamDTM_LT algorithm performs well in complex scenes such as target deformation,illumination change and partial occlusion.The tracking speed reaches 40.7fps,meeting the needs of real-time target tracking.