Quantitative analysis of asphalt road surface crack and pothole using fixed video data

To quantify asphalt road surface crack and pothole based on fixed video, this study applies deep learning-based image processing techniques for defect segmentation. An improved Unet-based semantic segmentation model was developed with optimisations in the backbone network, attention mechanism and loss function, enabling pixel-level segmentation of road distress including linear cracks and potholes. Experimental results demonstrate significant improvements in evaluation metrics such as mIoU and mPA, confirming the effectiveness of the model optimisation. Crack pixel accuracy improved from 47.2% to 76.9%, and pothole accuracy increased from 71.2% to 82.8%, enhancing the overall segmentation performance. After constructing a camera model and performing lens correction, a quantification model for pavement distress was developed by integrating with semantic segmentation results. Length and area quantifications of road surface distress, primarily cracks and potholes, were achieved. By counting the foreground pixels in defect segmentation images generated through the improved Unet model and applying the established camera model's projection ratio, the distress area was accurately calculated.