YOLO v8_CAT: Enhancing Small Object Detection in Traffic Light Recognition with Combined Attention Mechanism

This paper introduces YOLO v8_CAT, an advanced object detection model designed to improve the accuracy of small and challenging object detection in traffic light recognition tasks. Traditional object detection models, including earlier YOLO versions, often struggle with accurately detecting small objects and differentiating traffic light states (GREEN, RED, YELLOW, and OFF) due to limited feature refinement capabilities. YOLO v8_CAT enhances the baseline YOLOvS model by incorporating a Combined Attention Mechanism (CAT), which integrates channel and spatial attention paths to prioritize critical features. The channel attention path focuses on essential feature channels, while the spatial attention path emphasizes important regions, allowing YOLO v8_CAT to handle small and complex objects more effectively. Evaluated on the Bosch Small Traffic Lights Dataset, YOLO v8_CAT demonstrates significant performance improvements, achieving 84.5% accuracy for GREEN, 63.8% for RED, 21.5% for YELLOW, and 25.8% for OFF, outperforming YOLOv8 by notable margins. In particular, YOLO v8_CAT shows a 12.4% increase in accuracy for GREEN and an 8% improvement for RED detection. Additionally, YOLO v8_CAT maintains the high real-time processing speed of YOLOv8 at 1010 FPS. These results establish YOLO v8_CAT as a superior model for traffic light detection in autonomous driving, enhancing both detection precision and robustness without compromising computational efficiency.