YOLOv8-based visual perception method for guide robots

Abstract

To assist visually impaired individuals, this paper proposes a lightweight YOLOv8-based method for real-time detection of traffic signals, crosswalks, and blind roads. A self-built campus road dataset of 951 images (eight categories) is constructed, annotated with Labelme, and split into training (761) and validation (190) sets. The eight categories are: blind road, pedestrian crossing, pedestrian light green, pedestrian light red, traffic light green, traffic light red, traffic light yellow, and pedestrian light yellow. Trained for 300 epochs (640×640 input), the model achieves a mean average precision (mAP50) of 0.891 on the validation set, with 0.994 for blind road, 0.993 for pedestrian crossing, and 0.887–0.965 for traffic lights. With only 3.0M parameters and 105/16 FPS on GPU/CPU, it meets real-time requirements for guide robots. The dataset and code are available upon request.