Enhanced YOLOv10m Model Performance for Front Vehicle Light Signal Recognition

This paper presents an enhanced version of the YOLOv10m model for front vehicle light signal recognition, incorporating the SEAM (Separated and Enhancement Attention Mechanism) attention mechanism. The proposed model significantly improves the performance of vehicle light signal detection, achieving a precision of $\mathbf{9 8. 9 \%}$, a recall of $\mathbf{1 0 0 \%}$, and a mAP50 of 99.5%. Additionally, the model maintains a competitive mAP50-95 of $\mathbf{9 3. 8 \%}$, demonstrating its robustness across a range of intersection over union (IoU) thresholds. The model contains 20,414,248 parameters, with a computational cost of $\mathbf{6 9. 9}$ GFLOPs and a model size of $\mathbf{4 1. 4 ~ M B}$, making it efficient for deployment in real-time systems. The results indicate that the integration of SEAM contributes to improved detection accuracy and recall, providing a reliable solution for traffic safety applications, particularly in scenarios involving low visibility or dense traffic conditions.