Research on Road-Sign Detection Algorithms Based on Depth Network

Unmanned system research has attracted increasing academic and corporate attention in recent years. Traffic-sign and road detection systems play an important role in unmanned systems. However, current algorithms perform poorly in reading traffic signs from a long distance and cannot satisfy the real-time requirements of accurate and rapid detection. Therefore, Yolov5s-Swin model is proposed in this paper. First, the Swin Transformer module and convolution module are fused, the ResUnit module in the Yolov5s backbone network is improved, the Crswin module is proposed, and the PAFPN in the network is modified to strengthen the ability of the model to capture local feature information and improve network detection accuracy. The Swin Transformer module uses the moved-window attention function to segment the image into different windows of a certain size but cannot establish a connection between neighboring windows. Consequently, we designed a method for hyperbolic window attention. This approach updates the window partition method such that the window block for attention calculation at each pixel changes, which can improve the receptive field, increase the information extraction ability of the target, and solve the information loss caused by decreased resolution during image training. Experimental results show that the proposed improved Yolov5s-Swin model exhibits significantly better detection accuracy than existing models.