Traffic Signal Detection Algorithm Based on Dual-GAM Improved YOLO26 under Nighttime Conditions

Abstract

The environmental recognition of autonomous vehicles depends much on real-time object detection, and identifying small traffic signals in low-light conditions during the night is a difficult problem. The size of traffic signals is small and vulnerable to the disruption caused by street lights and the lights of vehicles, which requires high-quality feature representation capability of detectors. To overcome such a challenge, this paper presents the Global Attention Mechanism (GAM) to YOLO26 and suggests a better algorithm, YOLO26-GAM. Initially, YOLO26 is chosen as the base by comparative experiments between various YOLO-series models. Next, Grad-CAM is applied to view the basis on which the model makes decisions, and GAM modules are added at the H12 and H15 locations in the neck network. Cross-experiments and ablation experiments are done to test the effectiveness of the attention mechanism and the insertion positions. Experimental findings on the curated TN-TLD nighttime traffic light dataset demonstrate that YOLO26-GAM performs the best of all the experiments with a Precision of 0.964, Recall of 0.947, mAP50 of 0.966, mAP50-95 of 0.657, and F1 of 0.955. It is the only configuration that is superior to the baseline, without the attention mechanism, in all evaluation metrics, confirming the efficiency of the proposed method in detecting nighttime traffic signals.