The Research On Traffic Sign Detection Algorithm Based On Improved Faster R-CNN And YOLOX

With the continuous progress in the field of intelligent driving,the focus of automobile safety is also moving closer to active safety,and environmental perception technology has become an important topic that needs to be tackled urgently.Traffic sign detection is a key subdivision of intelligent vehicle environmental perception,researchers and auto manufacturers have never stopped researching and exploring it.This paper conducts research on the traffic sign detection algorithm based on improved Faster R-CNN and YOLOX.The main work is as follows:(1)Based on the Faster R-CNN algorithm,the backbone network VGG16 is replaced with Res Net50.The hybrid attention mechanism is integrated into the backbone residual structure.The multi-scale sliding window is used to improve the RPN network.Feature maps are generated in different depth convolution layers,and feature fusion is performed.Build a traffic sign dataset and use the K-means++algorithm to design anchor boxes.The results show that the improved algorithm improves the detection accuracy m AP from 85.99% to 94.38%.(2)Based on the YOLOXs algorithm,the shallow features of the backbone network are extracted and multi-feature fusion is performed.A shallow feature prediction head is added to form a multi-prediction head detection layer,and the lightweight ECA attention module is integrated into the Neck feature enhancement network to form an improved M-YOLOXs traffic sign detection algorithm.The improved algorithm improves the detection accuracy by 2.68%.(3)According to the CPU hardware environment,the M-YOLOXs algorithm is lightweight,using Mobile Netv3 and Ghost Net two lightweight networks to replace the backbone.Based on the depthwise convolution and pointwise convolution,the feature enhancement network is further lightened.Two lightweight traffic sign detection algorithms,LM-YOLOXs-m and LM-YOLOXs-g,are constructed.The number of parameters is reduced by more than 44%,and the weight file is lightened to less than 22 M.The effectiveness of the lightweight algorithm is verified from the aspects of accuracy and speed.