SODDA-YOLO: A Traffic Sign Target Detection Network in Complex Environment

In the field of autonomous driving, the recognition of traffic signs is a very important safety issue. However, traffic sign detection is easily affected by the environment. Complex climate changes and surrounding lighting changes will affect the accuracy of traffic sign detection. Therefore, we propose a traffic sign target detection network SODDA-YOLO (Small Object Detection with Deformable Attention YOLO) based on YOLOv8. In this network, we first propose a downsampling module combined with a self-attention mechanism, which enhances the model's ability to correlate global information by adding a self-attention mechanism to the c2f module. Secondly, when the model performs sampling operations, we use the HWD module to replace the original single convolutional module. When performing sampling operations on the feature map, Haar wavelet sampling is added to replace the direct use of the convolutional module to retain more detailed information of the feature map. Experiments show that the mAP of SODDA-YOLO on the dataset TT100K reaches 68.9%, which is 1.6% higher than that of the original YOLOv8.