Research On Single Target Tracking Based On Multi-feature Fusion

With the rapid development of computer and artificial intelligence technology,target tracking in the field of computer vision research has been a very important position,in real life and it can be seen in the study.To put it simply,target tracking is to obtain the motion trajectory of the target in a given video sequence,locate and track the target,and predict the position of the target in the next frame through the target position of the previous frame.Target tracking based on the Siamese network has reached the advanced performance,but it is still limited in semantic feature extraction.In this thesis,by analyzing the existing deep learning network model,and on this basis,a skeleton network based on pyramid feature network is proposed,which takes deep neural network as the backbone and integrates feature mapping at different levels.In addition,a new method is proposed to distinguish between positive and negative samples,to realize self-adjustment of loss function by adding loss terms,and to learn more discriminative embedding characteristics of target objects with similar semantics.In order to improve the tracking accuracy,a multi-feature fusion segmentation network was proposed by adding a mask branch.Through a large number of experiments and data analysis,the effectiveness and advantages of the proposed algorithm are verified.The main contents and contributions are as follows:(1)Aiming at the problem of insufficient feature extraction in shallow network,a Siamese network model based on pyramid feature network using Resnet50 as the backbone is proposed,it can integrate the feature map with strong semantic information in deep network and the feature map with weak semantic information but rich spatial location information in shallow network under the premise of less computation,and improve the feature utilization rate.Experiments show that it is better suitable for tracking small target objects.(2)Aiming at the asymmetry problem of RPN network,and analyzing RPN network,a new deep cross-correlation structure based on RPN network is proposed to classify and regress the feature graph,which solves the asymmetry and overall performance problem of RPN network training.At the same time,a discriminant instance embedding loss function for distinguishing similar objects is proposed to reduce the influence of interferences on targets.The expected average overlap(EAO)rate on VOT2016 was 2.9%higher than that on Siam RPN,and the accuracy rate on OTB2015 was 5.3% higher than that on Da Siam RPN.(3)Aiming at the effectiveness of the marquee problem,to improve the current tracking algorithm,multiple feature fusion mask segmentation algorithm is proposed,which can effectively track the target when the target is flipped and the shape changes,and improve the framing accuracy of the target and the efficiency of matching and tracking.By combining multi feature fusion segmentation algorithm with Siamese network architecture,we can get more accurate anchor frame and improve tracking efficiency.Compared with no segmentation algorithm,the expected average overlap rate on vot2016 is increased by 2.1%.