This paper addresses the challenges of mobile traffic sign detection in dynamic environments by proposing a lightweight model, RePCMA-YOLOv8n, to enhance feature extraction efficiency and detection performance. The RePCMA module integrates Partial Convolution (PConv), re-parameterization strategies, and multi-scale attention mechanisms (EMA), effectively reducing redundancy in feature extraction to lower the number of parameters and floating-point operations, thereby improving overall model performance. Specifically, to tackle the common issue of varying target sizes in traffic sign detection, this study introduces the Normalized Wasserstein Distance (NWD) as an optimization loss function, significantly enhancing the model’s detection accuracy for targets of different sizes. The model demonstrates exceptional adaptability across various scenarios. Experimental results show that on the LISA dataset, the algorithm achieved an mAP50 of 97.9%, and its outstanding performance was further validated on seven other public datasets using mAP50 and mAP50-90 metrics.