Research On Deep Learning Target Tracking Algorithm Based On Siamese Network Architecture

Target tracking is a challenging branch in the field of computer vision.With the rapid development of computer technology and artificial intelligence technology,the performance of target tracking algorithm is gradually improving.In the development of target tracking algorithm,target tracking algorithm based on siamese network has gradually become a research hotspot in the field of tracking because of its advantages of fast speed and high precision.However,at present,this kind of algorithm still has many defects,which limit its application.Firstly,the siamese tracking algorithm can’t automatically update the template during the tracking process,so that the template can’t provide enough target features in the late tracking.In addition,because the core of siamese tracking algorithm is the similarity calculation of template feature and search feature,the anti-interference ability of this algorithm is insufficient.Secondly,when the target moves fast and the motion is fuzzy,the tracking algorithm will fail due to insufficient features extraction.Finally,in practical scene applications,the tracked target in the field of view is usually not perpendicular to the horizontal axis of the observation surface,but with a certain angle,that is,the common target tracking algorithms can’t estimate the real pose of the target in the field of view.In order to solve the above problems,the Siamese-RPN(SiamRPN)algorithm is improved as follows:(1)In order to improve the feature expression ability and anti-interference ability of SiamRPN algorithm,a siamese attention module based on position attention module(PAM)and channel attention module(CAM)is introduced into the original algorithm.The attention module combines the advantages of position attention mechanism and channel attention mechanism,takes into account the context of local features and the interdependence between different channels of feature map,enhances the strong response of the tracked target,suppresses the weak response of surrounding interference items,and improves the anti-interference ability of the algorithm.In addition,the siamese attention module also integrates the characteristic graphs of two channels in the tracking process,which provides an implicit template update strategy for SiamRPN algorithm.(2)When the target has the problem of motion blur due to fast motion,the backbone network of SiamRPN can’t extract enough target feature information,resulting in the failure of the tracker and the problem of tracking drift.Therefore,this paper designs a target location network based on LiteFlowNet3 optical flow estimation network.When the tracker has the problem of tracking drift,it relocates the target through the target location network,and then uses the target location restart algorithm after relocation to improve the performance of the tracking algorithm.Because the core of the target location network uses the lightweight LiteFlowNet3,it has no great impact on the speed of the algorithm itself.In addition,the target location network proposed in this paper is an independent network.At the same time,in order to verify the effectiveness of the network in the tracking field,this paper makes ablation experiments on the network,combines the network with SiamFC and DaSiamRPN respectively,and tests the performance of the algorithm.After testing,the performance of the two algorithms is improved to varying degrees after combining the target location algorithm.(3)At present,most of the target frames output by tracking algorithms are vertical frames,which can’t reflect the rotation angle information of the target in the field of view,which is very important in some specific occasions.In order to solve the above problems,this paper designs an angle decision network based on reinforcement learning,and outputs the rotation angle of the target in the field of view through the angle decision network.Before the training,in order to increase the decision-making ability of the agent,this paper first modifies the data set and adds the angle information in the data set.In the training stage,this paper adopts the combination of supervised learning and reinforcement learning.The advantage of this method is that the network itself has sufficient feature extraction ability and the decision-making ability of the agent is also strong.Therefore,the network trained by this method is completely offline without online fine-tuning in the tracking process,which improves the speed of the algorithm.