Shape-Prior Driven YOLO Model for Pothole Detection to Assist Visually impaired pedestrians

Visually impaired pedestrians encounter substantial safety challenges during travel, often due to their inability to detect environmental hazards effectively. While voice-guided navigation technology has been developed to provide assistance, these systems remain chanllenging in detecting small potholes or uneven surfaces on sidewalks, which significantly impacts travel safety and convenience of Visually impaired pedestrians. To address this, we propose a YOLO framework integrated with shape prior, termed YOLO-SP, designed particularly for sidewalk pothole detection. Specifically, we propose a novel scheme that integrates shapeprior enhancement strategy into the essential YOLOv9 framework, focusing on the capability to capture the feature representation of pothole contours. Meanwhile, we constructed a private dataset to validate its effectiveness in the particular scene of sidewalk pothole. Experimental results show that YOLO-SP achieved the best performance with detection accuracies of 0.952 and 0.918 on public and private datasets, surpassing the state-of-the-art (SOTA) YOLOv9 by 1.0 and 5.3 percentage points. The proposed YOLO-SP demonstrates superior accuracy and recall, making it particularly valuable for enhancing the safety of Visually impaired pedestrians.