Abstract: To address the issue of failed traffic sign detection caused by factors such as small scale, high-speed movement, and adverse weather conditions, a traffic sign detection model based on coordinate convolution and optimal transport label assignment is proposed. In order to take into account the embedded platform deployment and detection speed, the proposed model builds upon the YOLOv5s detection model. Firstly, spatial information is perceived and the feature representation capability for small-scale traffic signs is enhanced by utilizing coordinate convolution with additional coordinate channels. Secondly, an optimal transport assignment method is employed to seek globally optimal object  label assignments, reducing the number of ambiguous bounding boxes and improving the utilization of training data. Finally, a SIoU loss function incorporating angle loss is utilized to enhance the con- vergence speed and detection capability of predicted bounding boxes. Experimental evaluations of the proposed model on CCTSDB and TSRD traffic sign datasets demonstrate significant improvements over the original YOLOv5s model. Compared with the YOLOv7 model, the proposed model achieves a 2.35% increase in mAP_0.5 and a 1.45% increase in mAP_0.5:0.95 on TSRD dataset, while performing on par with YOLOv7 on the CCTSDB dataset. Moreover, the proposed model exhibits more than 2.5 times faster detection speed than YOLOv7 on both datasets, highlighting its excellent detection accuracy and speed.

Key words: traffic sign, coordinate convolution, object detection, optimal transport assignment