Research On Traffic Marking Detection Of Intelligent Vehicles Based On Deep Learning

As one of the future directions of automobiles,intelligent vehicles and their key technologies are the focus of research areas for researchers.Environment perception is the most basic key technology of intelligent vehicles,which provides accurate road scene information for intelligent vehicles.Traffic marking detection is the key content of intelligent vehicle environment perception.There are many kinds of traffic markings,mainly including lane lines and ground traffic signs.In this thesis,lane lines and ground traffic signs are detected and studied based on improved U-Net and YOLOv5 s network algorithms.Based on the semantic segmentation method,the U-Net network structure is improved,and a lightweight two-branch lane line detection algorithm is proposed.The complexity of the algorithm is reduced and the real-time detection is improved by changing the backbone network and the convolution module of the decoder.In order to improve the detection accuracy of the algorithm,two modules,spatial pyramid pooling and attention gate,are introduced.The spatial pyramid pooling module performs multi-scale fusion of global and local features of deep semantics.The attention gate module trains the target on the shape of the lane lines and enhances the regional features of the lane lines.In order to improve the detection accuracy of the YOLOv5 s algorithm in complex road scenes,this thesis proposes a ground traffic sign detection algorithm based on the improved YOLOv5 s network.Firstly,the adaptive spatial feature fusion module learns the weight parameters and assigns them to the features of different scales,so as to realize the efficient fusion of features of different scales.The coordinate attention module is embedded in the backbone network of YOLOv5 s to improve the ability of the backbone network to extract precise location information and regions of interest.The EIo U loss function is used as the object localization loss,which makes the network converge faster and the localization effect better.This thesis verifies the improved algorithm based on the data set and intelligent vehicle platform.Compared with the U-Net algorithm,the FLOPs of the improved algorithm are reduced by 88.46%,the parameters are reduced by 43.20%,and the detection accuracy is improved by 1.52%.Compared with the YOLOv5 s algorithm,the overall m AP of the improved algorithm has increased by 2.5%.For the video captured by the camera of the intelligent vehicle platform,the FPS of the improved algorithm reaches 31 frames/s and 41.69 frames/s,respectively.The detection accuracy of lane lines is 90.02%,and the detection accuracy of ground traffic signs is 90.21%.The above results prove that the traffic marking detection algorithm proposed in this thesis based on improved U-Net and YOLOv5 s networks has strong practicability and can meet the basic requirements of intelligent vehicle traffic marking detection.