A Lightweight Deep Learning Model for Real-time Detection and Recognition of Traffic Signs Images Based on YOLOv5

Abstract

The rapid and accurate identification of various road traffic signs is an important research topic in automotive vision systems. Specifically, the correct identification of road signs is an urgent problem requiring effective solutions to facilitate automatic driving. This paper proposes a new approach, PP-LCNet-P2-CT, for the detection and recognition of urban road signs in an automotive vision system using an improved YOLOv5 deep learning model. The main improvement of the PP-LCNet-P2-CT model includes the following: (1) Replacing the YOLOv5 backbone network with the lightweight network PP-LCNet to improve the real-time performance of the detection network; (2) Adding a small target detection head to the detection head to meet the needs of target detection with different scales and mitigate the adverse effects caused by drastic target scale changes; and (3) Integrating the CBAM convolutional block attention model that focuses on target features and the transformer coding block that can capture different local information to ensure the accuracy of lightweight model target detection. The model was tested with the Tsinghua traffic sign dataset, TT100k. The results show that the mAP index of the PP-LCNet-P2-CT model is increased by 29.84% and the FPS is increased by 24.05%, while the number of model parameters is decreased by 32.78% and the GFLOPs decreased by 34.41% compared with the classic YOLOv5 algorithm. The PP-LCNet-P2-CT model allows complex deep learning to be used successfully for unmanned ground vehicles (UGVs) with ordinary computing speeds and high real-time requirements.