Research On Traffic Sign Detection Algorithm Based On YOLOv5 Network Architecture

Traffic sign detection technology has always been a hot spot and difficult point in the field of intelligent driving.YOLOv5 Network architecture has the advantages of small model and high accuracy in traffic sign target detection.However,the traffic sign detection task is often affected by the complex road environment and other problems,which leads to the increase of the difficulty of detection.Meanwhile,there exit problems in the existing data sets,such as uneven quality of traffic sign target features,incomplete model extraction feature information,unsensitivity to small target features and insufficient global feature processing ability.Such problems greatly reduce the performance of the detection model and fail to achieve the ideal detection effect.In order to solve the above problems,this thesis improves the YOLOv5 s target detection network architecture as follows:(1)Due to the diversity of YOLOv5 algorithm models and different model structures,two lightweight YOLOv5 s baseline models were selected to distinguish the two models between baseline A model with no slice operation and baseline B model with slice operation.In this thesis,we analyze the TT100 K traffic sign dataset in experiments based on two baseline models.(2)In view of the problem that the YOLOv5 s baseline A model fails to accurately pay attention to the feature information of different levels and cannot adaptively adjust the proportion of important information among the features,A traffic sign detection algorithm based on GAM-ASFF-YOLOv5 s is proposed.The Global Attention Mechanism module is added to the feature extraction part of the YOLOv5 s baseline A model to improve the utilization of feature information at different levels,so that the model can better extract small targets at different levels.In addition,the Adaptively Spatial Feature Fusion algorithm is introduced before the model detection head to adaptively adjust the feature fusion ratio between each feature layer,suppress the feature information conflict and retain the important feature information for detection.The experiment shows that the improved algorithm model effectively improves the detection accuracy.(3)In view of the problems of insufficient global feature information and insufficient feature fusion extracted from the baseline B model of YOLOv5 s in complex environment,a YOLOv5 s traffic sign detection algorithm integrating self-attention mechanism is proposed.The Swin-Transformer module based on the shift window is introduced into the feature extraction layer to increase the information interaction between feature images to obtain multi-scale image features;The feature fusion section introduces the ViT module and the Swin-Transformer module based on the multi-head self-attention mechanism to obtain the global feature information of the image to be measured;After that,the image splicing mode is empowered to prioritize the important feature information,and improve the detection efficiency of the model.The experimental results show that the improved baseline B model performs better.