Real-time traffic light detection for autonomous vehicles

With the rapid development of autonomous driving, the traffic light detection is crucial for vehicle decision-making in terms of safety and real-time performance. This paper presents an improved YOLOv10-based traffic light detection model, combining multi-module optimization with HD Map-assisted strategies to enhance accuracy and efficiency. The C2f-MLCA module strengthens multi-scale feature extraction and reduces redundant computations. The CCFM-Neck structure optimizes feature fusion to improve small object detection. A map-based ROI cropping strategy narrows the detection area, reducing computational costs. A dynamic switching strategy for long- and short-focus cameras addresses both near and far detection needs. Data augmentation is optimized to prevent orientation-induced classification errors. Experiments on the HSTLD and BSTLD datasets show that the model achieves an mAP50 of 80.3% on HSTLD (0.7% improvement) and 61.2% on BSTLD (0.8% improvement), with a 7.3% reduction in FLOPs. The model outperforms mainstream models in accuracy, inference speed, and size, demonstrating strong applicability in real-world autonomous driving scenarios.