UCN-YOLOv5: Traffic Sign Object Detection Algorithm Based on Deep Learning

Traffic sign detection plays an important role in traffic safety and traffic management. In view of the complex and changeable environment and detection accuracy of traffic sign detection, this paper proposes UCN-YOLOv5 model based on the framework of YOLOv5.This model first replaces a new backbone network, which uses the core module RSU of U2Net to enhance the feature extraction of the network. Then, ConvNeXt-V2 is integrated, and the C3 module of its Block and YOLOv5 network is used to construct the C3_CN2 structure. The utilization of the proposed lightweight receptive field attention module LPFAConv in the Head Section represents a potential enhancement for the extraction of receptive field features. Finally, for small targets in traffic signs, Normalized Wasserstein Distance (NWD), which is insensitive to targets of different scales, is added to calculate the position loss function to replace the IoU metric to a certain extent, which further improves the detection ability of our model for traffic signs. Experiments on the TT100K dataset show that UCNYOLOv5 has excellent detection performance. Compared with the baseline model (Y0Lov5s, YOLOV5m, YOLOV5l), it improves the Map.5 index by 5.9 %, 4.9 % and 4.6 %; in the Map.5:.95 index, it is 4.4 %, 3.5 % and 2.8 % better. Moreover, the enhanced algorithm demonstrated favorable performance on the LISA and CCTSDB2021 traffic sign datasets. This research has important value for the accurate detection of traffic sign detection, and has guiding significance for in-depth research in related fields.