Research And Application Of Vehicle Tracking System Based On Hierarchical Feature And Correlation Filtering

With the deployment and development of new-energy vehicle autonomous driving technology,vehicle tracking has become the focus and difficulty of environmental perception in autonomous driving technology,it is essential to improve the accuracy and robustness of vehicle tracking in complex traffic scenarios.However,the complex and changeable environmental factors such as illumination and weather,as well as the background clutter and occlusion in dense traffic flow,have brought many challenges to the vehicle tracking task in the traffic road environment.In view of the above problems,this paper improves the existing vehicle tracking algorithm.A neural network model with more target-background discrimination capabilities and a target re-tracking mechanism under occlusion conditions are proposed,which leads to improved vehicle tracking algorithm performance.The main research content and innovation results of this paper are as follows:(1)This paper proposes to the add channel attention mechanism to the feature extraction network.The convolutional neural network is used to extract the feature map of the vehicle target,and the weight of each channel in the feature map is recalibrated.The guidance algorithm model pays more attention to the feature channels with high discrimination,enhances the robustness of the expression of the target appearance feature,and solves the problems of motion blur and illumination variation in vehicle tracking.(2)This paper implements the siamese network model embedded in the relevant filter layer and interprets the relevant filter algorithm as a differentiable layer,which is tightly coupled to the siamese network model,which improves the accuracy of the similarity between the calculation target and the candidate region,which helps solve the problem of background clutter.The end-to-end and shallow network structure ensures the real-time performance of the tracking algorithm.(3)This paper designs the focal loss function.In order to cope with the problem of the imbalance in the number of positive and negative samples as well as difficult and easy samples when training the model in the tracking task,which ultimately leads to the problem of poor model generalization ability.The focal loss function designed in this paper makes the training of the model pay more attention to the positive samples and difficult samples that are less in number but are more valuable to the performance of the model.Make the model have a stronger generalization ability for difficult samples with similar targets and backgrounds.(4)This paper proposes a target occlusion detection and target re-tracking mechanism combined with Kalman filter algorithm.The motion information of the target vehicle is fully utilized.Not only the prior information of the target motion and the target appearance information are combined to improve the accuracy of position estimation,but also the target occlusion is detected in real time.When the tracking algorithm based on the appearance information of the target fails,the system motion equation predicts the target position to be used until the target reappears,which improves the robustness of the vehicle tracking algorithm.The algorithm proposed in this paper is tested on three challenging public data sets.The experimental results show that compared with the current advanced tracking algorithms,the tracking algorithm proposed in this paper has higher accuracy and robustness,and can better deal with various challenges in the traffic environment.