In order to solve the challenges of low accuracy, high model complexity, and suboptimal real-time responsiveness in small-target traffic sign detection, this research introduces an innovative lightweight detection model, CM-YOLOv5s. The model integrates the CSPCM module into the backbone and neck networks to substantially strengthen the feature extraction capability, employing an optimized C3_Faster module to reduce model complexity, adding a dedicated small-target detection head to improve the recognition effect, and innovatively introducing the NWD metric to optimize the CIoU loss function for addressing localization deviations. The experimental validation shows CM-YOLOv5s attains 84.5% detection accuracy while maintaining efficient inference, with model parameters and computational costs reduced by 20.4% and 0.6 GFLOPs respectively compared to the original YOLOv5s, exhibiting superior comprehensive performance advantages over existing methods.