Research On Traffic Sign Recognition Algorithm In Complex Scenes

Traffic sign recognition provides road condition information for autonomous driving systems,enabling the prevention of traffic accidents by accurately and quickly identifying traffic signs.In complex scenarios such as low light conditions,occlusion,and sign deformation,the recognition of small traffic signs is greatly hindered,posing a challenge to quickly and accurately identify them.To address the issue of traffic sign recognition in complex scenarios,this paper conducts research on the traffic sign recognition algorithm based on YOLOv5,with the following main research contents:1)In complex and dynamic traffic scenarios,this paper proposes a traffic sign reco-gnition algorithm based on complex scene analysis.The improved algorithm utilizes the GIOU,which solves the issue of zero loss when prediction boxes overlap,thereby improving the accuracy of traffic sign recognition when targets overlap.By replacing the weighted non-maximum suppression module with the Cluster NMS module and using weighted averaging in the Cluster NMS module,the improved algorithm enhances the speed and accuracy of traffic sign recognition.The sign replacement technique,which does not compromise background features,is employed for dataset augmentation.In complex scenarios involving low light,occlusion,distortion,etc.The improved algorithm achieves accurate identification of traffic signs.For the recognition of small target traffic signs,the m AP@0.5 of this algorithm reaches 0.912,improving the precision by 0.299 compared to the original algorithm.2)In complex scenarios,where lightweight recognition algorithms suffer from low precision and slow speed,this paper presents a lightweight real-time traffic sign recognition algorithm.Building upon the lightweight algorithm YOLOv5n6,the improved algorithm incorporates the CA mechanism into the backbone network,strengthening the positional information of targets and enabling more accurate localization.The Si LU activation function in the SPPF module is replaced with Re LU activation function to create the sim SPPF module,which operates faster than the SPPF module.Data augmentation using the copy-paste technique increases the quantity of small targets,improving the precision of small target recognition.The m AP@0.5 of this algorithm reaches 0.91,achieving a precision improvement of 0.247 compared to the YOLOv8 algorithm.With a recognition speed FPS exceeding 30,the algorithm meets the real-time requirement.The FPS of this algorithm is 57.47,satisfying the real-time constraint.The algorithm’s GFLOPs is 4.4,and the model size is only 7.2MB.Compared to the lightweight algorithm YOLOv3-Mobile Netv2,this algorithm improves the precision by 0.06,increases the FPS by 14.27,and reduces the model size by25.3MB.