Comparative Study of YOLOv11-Segmentation Models for Vehicle Light Status Detection

Vehicle lights such as headlights, brake lights, and turn signals are critical for road safety but often go undetected when malfunctioning. This study explores real-time detection of these lights, including their ON/OFF status, using three YOLOv11 segmentation variants: yolo11n-seg, yolo11sseg, and yolo11m-seg. Models were evaluated on a labeled test dataset and CCTV footage from a parking lot. YOLOv11-m achieved the highest accuracy (box mAP@50: 0.682, mask mAP@50-95: 0.378), while YOLOv11-n offered the fastest and relatively stable inference speed (16.80 FPS). Challenges included differentiating between visually similar light elements and dealing with the limitations of fixed top-down CCTV camera angles, which reduced visibility. While YOLOv11-m was the most contextually accurate, YOLOv11-n proved suitable for real-time use in low-resource settings. The results highlight trade-offs between accuracy and efficiency and suggest future improvements using temporal filtering, environmentspecific tuning, and post-processing to enhance deployment in surveillance-based vehicle inspection systems.