Research On Traffic Sign Detection Algorithm Based On Deep Learning

The role of traffic signs is to provide accurate road information to drivers and ensure the smooth passage of the road.Traffic sign detection is an important part of intelligent vehicle assisted driving and automatic driving.Because the categories of traffic signs are different,and multiple targets may overlap and block in the detection process,how to detect traffic signs accurately and efficiently is a challenge.Target detection methods can be divided into traditional methods and methods based on deep learning.Traditional methods extract features manually and then match templates,which have the problems of low detection accuracy and poor real-time performance.Deep learning methods often have good detection accuracy and good detection efficiency.These algorithms have been widely applied to various types of automatic driving systems.Therefore,this paper is based on the national key research and development plan sub topic “autonomous electric vehicle environment perception technology” project as the main support,this paper studies the traffic sign target detection algorithm based on the existing deep learning target detection algorithm,aiming at improving the detection accuracy and speed of the algorithm.(1)Based on the in-depth analysis of the single-stage representative algorithm YOLOv3,this paper proposes an improved traffic sign detection algorithm based on YOLOv3.The basic idea is that firstly,the feature fusion module is redesigned,and a multi-scale traffic sign target detection method with five detection scales is designed,which improves the detection effect of small targets;Secondly,the improved Kmeans++ algorithm is used to recluster anchor frames;Finally,GIoU is used to replace the original MSE loss function in YOLOv3 to improve the positioning accuracy.The result shows that the algorithm improves the detection accuracy while ensuring the realtime performance.(2)In order to further improve the detection accuracy of the traffic sign algorithm,this paper makes an improvement based on the in-depth analysis of the two-stage representative algorithm Faster R-CNN algorithm.Firstly,the FPN is constructed to obtain more semantic information;Secondly,ROIAlign is used to avoid the accuracy loss caused by pooling operation;Finally,deformable convolution is added to the backbone network to enhance the flexibility of network sampling.At the same time,in order to improve the speed of the model in detection and inference,TensorRT model is further used to accelerate the model.The experimental results show that: compared with the original Faster R-CNN,the accuracy of this algorithm is greatly improved,and the inference time of TensorRT model is effectively reduced,which meets the requirements of real-time detection.This paper trains the improved traffic sign detection model based on YOLOv3 and Faster R-CNN,designs the visual operation interface of the detection algorithm,and carries out the real vehicle experiment of the improved algorithm on the intelligent driving platform.The source code is completed by Python3.7 and C++.The deep learning framework of model training and testing is Pytorch.The operating system is Ubuntu 18.04,and the development tool is Pycharm.Real vehicle experiments show that the improved algorithm can realize the traffic sign target detection in real traffic scenes,and the algorithm has good engineering application value.