Research On Lightweight Network Detection For Road Traffic Targets

Traffic target detection is one of the important research issues of autonomous driving technology in the field of computer vision.During the autonomous driving process,the traffic target can be divided into: target detection objects represented by pedestrians,vehicles and traffic signs;semantic target objects represented by lane lines.At present,deep learning algorithms have been widely used in the field of detection and identification of traffic targets,which is the hotspot of the current research.Some new lightweight detection networks have obtained more and more applications in autonomous driving systems due to fast running speed and high detection accuracy.However,with the continuous improvement of target detection accuracy requirements,people are constantly improving these networks to improve the accuracy,speed,and adaptability of different detection targets.In this article,the detection needs of the warning system and lane departure of the alarm system are deviated.Several existing networks have been carried out on the problem of the target detection of road traffic signs,vehicles and pedestrians,and the issue of the segmentation of road lanes to provide reference for supplies.The main work is as follows:(1)For the small target detection tasks of the traffic signs,the traffic signs based on the first stage of YOLOV5 S are proposed to identify the traffic signs.This method is under the Ghost Net model of 3 × 3 computing cores.Bottleneck module,replacing all the Bottleneck modules in the network,combined with the cross-stage local network(CSP-Net)module to generate Ghost Bottleneck CSP module;and insert an ECA module in the three output terminals of Backbone into an ECA module.,Enhance the network Expression of key information.Moderate with the new YOLOV5-GHOSTECA network.After testing,the MAP value of the network on the TT100 K traffic logo dataset was 0.873,which improved the detection accuracy of the algorithm.The FPS was 67,retaining the real-time nature of the algorithm.(2)Aiming at the detection task of vehicles and pedestrians,a YOLOv4-tiny-based pedestrian and vehicle object detection algorithm is proposed,and the Convolutional Block Attention Module(CBAM)is introduced after the CSP Darknet53-tin module,so as to enhance the extraction of key feature information by the network;At the same time,the bidirectional Feature Pyramid Network(Bi FPN)is used to replace the Feature Pyramid Network(FPN)of the original network to improve the feature fusion capability of the network.After testing,the m AP value of the new network on the BDD100 K dataset is 82.1%,which is 2.2% higher than that of the original network,and the FPS is 85,which meets the real-time detection requirements of vehicles and pedestrians.(3)In lane line detection,aiming at the problem of low detection accuracy caused by insufficient feature extraction by Ultra-Fast Structure-aware Deep Lane Detection(UFLD),it is proposed to add convolution and pooling operations before the fourth Bottleneck module in the original backbone network of Mobilie Net V2 to improve the feature extraction capability of the network.The last three bottleneck modules in the original backbone network of Mobilie Net V2 are removed to avoid the phenomenon that the network training gradient disappears due to the network being too deep.Then,the Res Block backbone network in the UFLD network is replaced with the improved Mobilie Net V2 backbone network to form the Mobilie Net V2-UFLD network.Finally,the Tusimple2017 dataset is used for experimental verification,and the experimental data show that the new network has made a breakthrough in detection accuracy,with an accuracy rate of 92.1% and an FPS of 125,which fully meets the real-time requirements.