Real-Time Traffic Sign Detection Based on Weighted Attention and Model Refinement

With the rapid development of technology and artificial intelligence in today’s era, people are increasingly inclined to the technology of automatic driving, Real-time detection of traffic signs while driving is required for safety reasons. In this work, we improve the current state-of-the-art detection algorithm YOLOv7 as a basis and propose a lightweight traffic sign detection algorithm. First, we highlight the weights of salient features through a weighted attention module to make the network more focused on the target features. Then, we propose a lightweight convolution-based spatial pyramidal pooling fusion module (DW-SPPF) and replace the conventional convolution of the detection head with a lightweight depth-separable convolution to reduce the complexity of the model while maintaining detection accuracy. Finally, for the bounding box regression, we consider the directional relationship between the groundtruth and predicted boxes and improve the regression loss function to enhance the convergence speed of the loss. The experimental results show that the accuracy of the model in this paper is improved by 6.7% compared to the base network, and the detection speed is accelerated by about 51 frames/s through model refinement, which basically satisfies the real-time accurate detection of traffic signs.