The Research On Traffic Sign Detection In The Automatic Driving Systems

With the successive breakthroughs in many technological fields such as computing chips and deep learning algorithms,traffic sign detection algorithms have made great strides in recent years.However,in the face of the complex and changing traffic scenarios,the current traffic sign detection algorithm still has great room for improvement.This article proposes a series of methods to solve the current difficulties encountered in traffic sign detection.The main contributions of this paper are as follows.(1)This article proposes a traffic sign data composition method by observing the various changes of traffic signs in the actual traffic scenario,which effectively solves the problem of serious data imbalance and sample missing in the existing traffic sign data set.In this article,the feasibility as well as the robustness of the method is verified by experiments on the TT100 K data set.(2)The YOLOv3 algorithm has poor performance in small object detection.In this article,the shallow and deep features of the fusion network are used to further improve the size of the feature map to improve the detection capability of the YOLOv3 algorithm for small-sized objects,in addition,this article fuses the attention mechanism into the YOLOv3 network to more effectively capture the location of objects in complex backgrounds and improve the model performance.Experiments on multiple datasets have shown that the performance of the improved network presented in this article is a large improvement over the original YOLOv3 network,especially for small size objects.Finally,based on the above-mentioned improved network and the proposed traffic sign data composition method,a two-level traffic sign detection framework with the ability to detect and identify all traffic signs on the road is proposed.Experiments on the TT100 K dataset show the robust performance and strong scalability of the framework.Compared to other traffic sign detection algorithms,the traffic sign detection algorithm in this article has a large advantage in terms of both the number of classifiable categories and accuracy.