Research And Implementation Of Vehicle Target Detection Algorithm For High-density Traffic Scenes

Vehicle detection in high-density scenes is one of the hot research contents in the field of computational vision,which is of great significance for the research and development of vehicle-road cooperative perception.When performing vehicle target detection based on traffic scenes,it is often easy to produce poor target recognition efficiency and slow detection speed due to the dense distribution of vehicles in the scene or the distance of the target to be detected from the acquisition device.Among deep learning algorithms,target detection for computer vision tasks has become mainstream.Since the traditional anchor-frame-based target detection algorithm is inefficient and cannot guarantee the balance of accuracy and speed,this paper optimizes the existing deep learning algorithm based on the anchorless detection method to achieve efficient and high-precision dense vehicle target detection.The main work of the paper is as follows:(1)A fully convolutional single-stage detection anchor-free detection method is proposed to eliminate the predefined Anchor box and avoid the complex computation associated with the anchor box.The center point of the target vehicle is directly computed to obtain the predicted target position,which improves the efficiency of the target detection task and greatly increases the detection speed.(2)In the algorithm framework,this paper designs a vehicle target detection algorithm based on Retina Net algorithm for high-density traffic scenes,and optimizes the algorithm in terms of the prediction method and the quality estimation of the prediction results so as to obtain performance improvement.The feature pyramid structure FPN is proposed to predict targets at different scales using layers,and the structure is optimized to add a prediction process to the layer responsible for the detection of small targets at long distances,thus reducing the probability of overlapping detection areas and improving detection accuracy.The quality of the output prediction frame is estimated at the detection layer,and the intersection and merging ratio(Io U)is calculated to improve the regression accuracy.The loss functions are set separately for different detection tasks,and the loss value cases of this algorithm are obtained by introducing the loss calculation of Focal loss,combined with Io U loss and binary crossentropy loss function.(3)The data of the traffic congestion scenes on the South Second Ring Road in Beilin District of Xi’an are collected,and their own datasets are produced by video key frame filtering and 2d image labeling,and output as COCO dataset format for testing the algorithm of this thesis.The generated dataset is used to implement this paper’s algorithm and compare it with YOLOv3 and Retina Net algorithms for experiments.Finally,in order to better demonstrate the research results of this paper,the system integration of the algorithm model studied in this paper is based on C/S architecture,and a dense vehicle target detection platform is developed and implemented.From the analysis of experimental results,compared with Retina Net algorithm,the algorithm of this thesis has been greatly improved in terms of detection speed fps index and target detection accuracy.The detection speed is increased from 32 fps to 52 fps,and the detection accuracy of vehicles is increased from 52% to 90%,while the detection capability of targets in high-density scenes is improved.