Highway foreground picture distress detection method based on homography matrix and object detection

To address complex equipment installation, limited detection accuracy, and the lack of standardized distress quantification in highway pavement inspection, this study proposes a foreground distress detection framework integrating homography estimation and object detection. Road marking features are used to estimate a homography matrix for perspective correction and image-to-ground coordinate transformation, enabling distress size measurement while simplifying rear-mounted equipment configuration. A dynamic dataset partitioning strategy is developed to alleviate small-sample and class-imbalance problems, reducing the average category distribution deviation to 0.14% and improving precision, recall, F1-score, mAP50, and mAP50–95 by 1.69%, 6.29%, 4.40%, 7.42%, and 7.76%, respectively. A cross-scale information fusion module further enhances multi-scale feature representation, yielding mAP50 gains of 2.85%, 2.00%, and 3.11% for YOLOv8, YOLOv11, and DETR, respectively. Finally, a heatmap-based quantification method uses multi-layer activation responses to estimate distress area. At a 30% abnormal-pixel threshold, the method achieves the best performance, with an average area estimation error of only 4.3%. These results demonstrate that the proposed framework improves both pavement distress detection and standardized geometric quantification.