Research On Target Tracking Algorithm Based On Siamese Network Structure

Object tracking is one of the important branches of computer vision,The main task of object tracking is to achieve accurate track under different challenge factors.Under the influence of illumination change,scale change,fast movement and complex background interference,there is room for improvement in the robustness and accuracy of object tracking algorithm.Based on fully-convolutional siamese networks for object tracking,the paper aims to improving the accuracy of the algorithm and the ability to identify and locate the target.The main research work is as follows:(1)Aiming at the problem that the tracking performance of the object tracking algorithm of full convolutional siamese network(SiamFC)is reduced under complex background,especially when it is disturbed by similar background,a object tracking algorithm of siamese network combining attention mechanism is proposed.First,add the channel attention module to the example branch to make the target get more attention.According to the weight coefficient of the channel attention module feature vector and the attention of the target information,the channel attention feature map is obtained,and the channel attention feature map inter-correlation operation of graph and search branch.Secondly,when the network model is training,in order to fully extract the characteristics of the target positive and negative samples,a triple loss function of the relationship between the input samples and positive and negative instances is constructed.Finally,Euclidean distance is used to calculate the pixel distance of the output feature graph in space,and the highest point of the response value is obtained as the object position.Experiments show that the accuracy and success rate of OTB2015 datasets are improved by 3.8% and 9.8% compared with SiamFC,and the accuracy of VOT2016 datasets is improved by 3.6% and the average expected overlap rate is 33.1%,indicating stronger robustness.Qualitative and quantitative comparison with other algorithms shows that the algorithm has better tracking performance when dealing with complex background.(2)Aiming at the problem that the tracking performance of Full Convolutional Siamese Network(SiamFC)algorithm decreases under complex background,especially when scale changes,a deep siamese network object tracking algorithm based on per-pixel classification and regression is proposed.Firstly,the Alex Net is replaced with the Res Net50 network which can extract more deep semantic information,and the influence of deep network filling operation is eliminated by using spatial aware sampling strategy.Secondly,multi-layer feature fusion is introduced to realize the information fusion of different network layers,and the last three residual block feature graphs are connected in series.Finally,the classification of the unanchored frame and the offset of the distance between the points of the regression frame and the four edges of the real bounding frame are introduced to determine the width and height of the prediction frame.Experimental results show that the accuracy and success rate of OTB2015 datasets are 7.6% higher than that of SiamFC and 13% higher than that of SiamFC.Compared with other algorithms,it is shown that this algorithm performs well in dealing with complex background.