Research On Vehicle Multi-target Tracking And Detection Based On Anchor Free Model And Attention Mechanism

Multi-target tracking is a key technology for the development of intelligent transportation and intelligent driving.Analyzing the number of vehicle targets obtained by tracking and the trajectory of vehicles can effectively improve the comprehensive management level of road monitoring,traffic flow statistics,assisted driving and driving habits analysis.In recent years,deep learning has promoted the development of computer vision.Scholars have proposed a large number of multi-target tracking methods based on deep neural networks.This type of method correlates the target detection results obtained by the network frame by frame,and realizes multi-target tracking based on detection,which improves the performance of the tracking method.However,vehicle targets have different shapes,high moving speeds,and severe mutual occlusion,which leads to problems such as difficulty in feature extraction in the detection process of vehicle tracking and increased detection calculations.This article starts with the detection process of tracking,and handles multi-target tracking tasks with new target detection forms and feature extraction processes,which not only improves the efficiency of the tracking method,but also enhances the accuracy of the tracking method.First,this paper analyzes and studies the principles of Anchor Free detection network and joint learning model,and proposes an Anchor Free joint model vehicle multi-target tracking method.This method introduces a single-network multi-task learning joint model into vehicle multi-target tracking,and realizes target detection and re-identification feature extraction with a single-network multi-branch structure.The deep feature aggregation network improved by variable convolution is used as the feature extraction network of the model to obtain the feature map of the video frame and improve the adaptability of the network to vehicle targets.The parallel branch structure is used to realize Anchor Free target detection and re-identification feature learning,reducing repetitive calculations and ensuring the detection level of the model.Because the method only uses a dense computing network to complete the main calculation process of the tracking task,the efficiency of vehicle multi-target tracking is improved.Secondly,this paper analyzes the feature extraction network of vehicle multi-target tracking,and studies the Attention mechanism.It is found that the Attention mechanism can be used to strengthen the network model’s ability to acquire features of vehicle targets.Therefore,this paper proposes a vehicle multi-target tracking method combined with Attention mechanism.This method combines the feature enhancement network combined with the Attention mechanism with the original feature extraction network.Under the premise of ensuring the efficiency of tracking and detection,the feature performance of the vehicle target on the feature map is improved,so that the feature map has a strong correlation with the target.Reduce the ID change between different vehicles in the vehicle multi-target tracking,the occurrence of tracking trajectory loss,etc.,and improve the accuracy of the vehicle multi-target tracking.Finally,this paper conducts experiments on the two proposed vehicle multi-target tracking methods based on improved tracking and detection,and compares them with the existing methods.The KITTI-Tracking data set and UA-DETRAC data set are used to train and test the method in this paper.The experimental results show that the vehicle multi-target tracking method proposed in this paper is better than the existing methods on the premise of ensuring the efficiency of reasoning.This effectively reduces the occurrence of vehicle target misdetection and ID switching problems,and improves the robustness of the vehicle multi-target tracking method to a certain extent.