Research On Target Tracking Algorithm Based On Deep Learning

Target tracking is a hot research direction in the field of computer vision,showing great practical significance and broad application prospects in intelligent security,automatic driving,etc.At present,the target tracking method still has limitations in dealing with the problems such as dramatic changes in the appearance of the target,the disappearance and recurrence of the target,and the limited computational resources.In this paper,deep learning method is used to explore target tracking.The main work and innovation are as follows :1.Aiming at the problem that the sharp change of target appearance in complex scenes seriously affects the tracking accuracy,a siamese network target tracking algorithm based on dynamic template update(DTU-Siam)is proposed in this paper.The algorithm uses the modified Resnet-50 as the backbone network to extract depth features and enhance the feature extraction ability of the network.For the response map obtained by cross-correlation operation,the position and proportion of the target object are predicted directly by anchor free method.For template updating,the template dynamic updating strategy is used to determine whether the template is updated.If updating is needed,the template updating subnet is used to estimate the best template for the tracking of next frame.Experiments on OTB2015 and VOT2018 public datasets show that the DTU-Siam proposed in this paper effectively improves the accuracy and stability of tracking compared with the mainstream algorithms.2.Aiming at the problem of target disappearance and recurrence caused by complex situations such as target occlusion and out of view in long-term target tracking scene,a siamese network long-term target tracking algorithm based on meta-update and re-detection(Musiam-LT)is proposed.Firstly,for the local tracker of the twin network,a template updating mechanism based on meta-learning is designed by introducing the long short-term memory network to update the target template when the target does not disappear.Secondly,for the target disappearance,the template guided multi-scale re-detector is used to obtain the proposal region of the tracking target from the global image,and then it is sent to the local tracker for target recurrence judgment and location tracking.In the public datasets Lasot and VOT2018＿LT,the Musiam-LT algorithm proposed in this paper can effectively improve the long-term tracking performance.3.Aiming at the practical application scenario under limited computing resources,this paper uses Rk3399 Pro domestic embedded platform to implement the target tracking algorithm.Combined with the characteristics of chip hardware,the siamese network target tracking process and embedded deployment difficulties are analyzed in this paper.The backbone network,cross-correlation operation and classification regression sub-network are optimized,and the algorithm process is multi-threaded from the engineering point of view.Through the actual application scenario test,the embedded target tracking application deployed in this paper can stably track conventional targets at a near real-time speed.