Research On Video Object Tracking And Segmentation Algorithm Based On Siamese Network

The object tracker needs to predict the exact position and shape change of the object in real time in the video sequence,but due to the deformation of object itself and the influence of environment factors such as distractors,robustness and location accuracy of the tracker will often fell sharply,to establish a high robustness,high location accuracy of visual tracking is still faced with enormous challenges,the end-to-end video object tracking and segmentation framework and the tracker with high robustness are explored in this paper.The main work contents are as follows:(1)This paper proposes a high robustness tracker,namely an Anchor-Free Siamese Tracking Algorithm With Searching Center Point(AFST).At present,the siamese network tracking algorithms all lack the use of multi-stage features.In this paper,a multi-level feature fusion module is proposed to fuse multiple features of different layers to enhance the robustness and discrimination of features.In order to realize the balance of positive and negative samples and the mining of difficult samples,a unified sampling strategy is proposed.In order to simplify the structure of the tracking network and eliminate the false positive output of classification branches,this paper proposes a one-time prediction method based on pixel points.Through the verification of the experiment,the algorithm proposed in this paper has obtained advanced performance in robustness and positioning accuracy.(2)This paper proposes a unified end-to-end object tracking and segmentation framework based on siamese networks,namely an Object-Aware Tracking and Segmentation Algorithm(OATS).OATS is composed of a object tracking network and a segmentation network,which can track and segment the target in the video in real time with high positioning accuracy.In order to make the segmentation network segment only the tracking target,this paper uses the score of the classification branch in the tracking network and several cross correlation vectors to form a spatial constraint vector with high robustness and high precision.Then the spatial constraint vector is taken as the input feature of the segmenting network,and the refined mask is output after gradually merging the underlying features.Finally,using the AFST network to replace the tracking network in OATS and adjust the space constraint vector,forming the OATS+ network.The experimental results show that OATS+ has obtained high segmentation accuracy.