Siamese Network Tracking Algorithm Based On Dual Attention Similarity And Depth Map

Visual object tracking is an important problem in the field of computer vision.Today,it is widely used in many applications such as video surveillance,self-driving,sports event broadcasting,and robots.However,changes of the surrounding environment of the target,changes of the target itself,the change of camera angle and distance may all affect the tracking performance of the tracker.Most object tracking algorithms only process RGB images,and use cross-correlation to match image features,which makes it difficult for trackers to deal with occlusion,deformation and extreme environmental factors.This thesis studies the problem of appeal.Firstly,aiming at the problem that the existing object tracking algorithms cann’t deal with the occlusion and deformation of the target,a siamese network tracking algorithm based on the dual attention similarity is proposed.The method uses the siamese network to extract image features,calculates the pixel-level similarity between template features and search region features through the dual attention similarity matching module,and performs classification and regression operations on the similarity map.This enables improved performance of the method for occlusion and deformation problems.Secondly,aiming at the problem that the RGB tracking algorithms cann’t deal with the occlusion of the target and the dark light,an object tracking algorithm based on depth information redetection is proposed.This method uses a combination of RGB and depth images,predict the target tracking result through the RGB tracker,and the similarity of depth information is used to judge whether the predicted result is reliable or not.If the prediction result of the current image frame is unreliable,a new prediction result is generated through the occlusion recovery mechanism to improve the performance of the tracking algorithm.Finally,this thesis implements all models using the Pytorch framework.The ablation experiment and contrast experiment of siamese network tracking algorithm based on dual attention similarity matching and contrast experiment of object tracking algorithm based on depth map redetection are carried out by using public datasets.