A lightweight algorithm for detecting traffic signs

Abstract

In the task of traffic sign detection, traffic sign targets are of small scale and occupy a small proportion of the image. At the same time, the background is complex. If in abnormal weather conditions (such as rain, fog, etc.), it will seriously restrict the improvement of detection accuracy. Moreover, as the detection accuracy increases, the algorithms tend to become more complex and the model size increases. Aiming at the dual constraints of low detection accuracy and large model size, this paper has carried out lightweight model improvement based on YOLOv8n from the YOLO (You Only Look Once) series. Firstly, some C2f modules in the network are replaced with C2f_FasterBlock modules to simplify the complexity of the model. Then, since most traffic signs are small; to fulfill the needs of small target detection, the multi-scale detection structure is reconstructed, and P2 detection layer and detection head are added to retain the details of small targets; at the same time, the redundant P5 detection layer and detection head are pruned; thereby achieving the collaborative improvement of model lightweighting and detection accuracy. Experimental results show that the improved model achieves a precision (P) of 87.1% on the CCTSDB 2021 dataset, a recall (R) of 71.4%, mAP50 of 79.7%, and mAP50-95 of 49.9%. Compared with the original YOLOv8n, it has decreased by 0.2%, increased by 4.1%, 4.2%, and 3.2% respectively; meanwhile, the model size has been reduced to 3.7MB, meeting the lightweight deployment requirements of vehicle-mounted edge devices.