Research On Target Tracking Method Based On Siamese Network Architecture

Target tracking is one of the hotspots in the field of computer vision.It has a wide range of application prospects in human-computer interaction,military reconnaissance,unmanned driving and security monitoring.In recent years,the performance of tracking algorithm has been widely improved,but in the actual tracking environment,there are still many challenges,such as the target will be affected by background clutter,scale changes,occlusion and other complex situations.Based on the Siamese network tracking framework,this thesis studies and discusses the background clutter,scale change and occlusion of the target.The main work and contribution of this thesis are summarized as follows:(1)A Siamese network tracking algorithm based on attention mechanism is studied.Firstly,the template image and search image are trained by two branches,and then the more robust features are extracted by convolution neural network,which makes the target get better tracking results under the condition of illumination change and background interference.Through the analysis of different levels of convolutional neural network feature map,we find that the lower level feature detail information is more,the accuracy of the target is more robust,the higher level feature semantic information is more abundant,and the classification of the target is more robust.In order to get more robust target tracking results,this thesis uses different layers of features for fusion,and in the process of feature extraction,the attention mechanism is used to recalibrate the feature map,and the Ada Boost algorithm is used to weighted fusion the target feature map,which improves the reliability of the response map.In this thesis,the concept module is also used.On the one hand,it increases the width of the network and the adaptability of the twin network to the scale.On the other hand,it reduces the parameters and improves the speed of network training.(2)A Siamese network tracking algorithm based on correlation filtering is studied.Firstly,the correlation filter is used in the template branch to speed up the extraction of network features.At the same time,the scale estimation module is introduced to scale the target at multiple scales when detecting the target position.In the process of scaling the target,this thesis uses the region based fast HOG feature extraction algorithm to speed up the extraction process,so that the Siamese network can quickly extract the hog features,thus speeding up the tracking speed.The proposed algorithm not only makes full use of the advantages of the real-time correlation filter,but also combines the features of different layers,and makes the target keep better robustness when the scale changes.(3)A Siamese network tracking algorithm based on optical flow information is studied.First,the current frame and the previous frame of the image are passed through the optical flow network,and the position of the target is predicted by the optical flow information,because the optical flow method can detect the target motion information in the image,and then pass through ROI pool layer can transform the features of the candidate region of the target into a small feature map with fixed scale,and then form a search region through clipping,which matches the first frame template image,and finally obtains the predicted position and size of the target.The proposed algorithm not only makes full use of the optical flow network to extract the motion information of the target,improve the robustness of the target,but also takes advantage of the real-time performance of the correlation filter,and combines the characteristics of different layers,so that the target can maintain a better robustness and real-time performance.In this thesis,the above algorithm is tested on OTB 2015.The experimental results show that the Siamese network tracking algorithm based on layered attention mechanism can effectively solve the impact of background clutter and improve the accuracy of tracking.The Siamese network tracking algorithm based on correlation filter uses correlation filter layer to speed up the tracking speed,and a fast scale estimation module is added to the algorithm,which can effectively improve the robustness and real-time performance of the target in scale change.The Siamese network tracking algorithm based on optical flow information can solve the problem of the target under occlusion.By using optical flow network to extract the motion information of the target,the target can maintain a relatively high accuracy under occlusion.