Lightweight traffic sign detection based on an improved YOLOv7-Tiny

Accurate and fast recognition of road traffic signs is essential for autonomous driving systems to validate navigation information and ensure safe and reliable decision-making. Although current navigation systems often already provide expected information about sign locations, visual perception remains an important means of verifying the actual presence of signs. However, the limited computational power of in-vehicle devices limits the complexity of detection algorithms. To address this issue, we choose Yolov7-Tiny as an optimization baseline to improve detection speed and accuracy within hardware constraints. We propose a lightweight adaptive channel attention fusion convolution (EPConv) to reconfigure the efficient layer aggregation network structure to improve accuracy while reducing complexity. We introduce an improved spatial pyramid module (SPPCSP-ED) to enhance feature extraction and design a lightweight dynamic fusion detection head (LDFDH) to improve information exchange between classification and regression branches. In addition, the adaptive loss function (NCIOU) was formulated to reduce the sensitivity of the traditional loss function to small target localization. Experimental results on the CCTSDB2021 dataset show a 30% reduction in parameters, a 27% reduction in computational cost, a 36 FPS improvement in inference speed, and a 4.4% improvement in detection accuracy. Further evaluation on the VOC2007 and TT100K datasets confirms the strong generalization capability. Real-world vehicle testing verifies that the improved model can be effectively deployed in vehicles to meet real-time detection requirements.