Road Rockfall Detection by Integrating Feature Engineering with YOLO and Cascade Decision Fusion

Abstract

In roadside surveillance imagery, shadows, vegetation, vehicles, exposed pavement, and water stains may exhibit local textures and morphological characteristics similar to those of rockfalls, causing a standalone You Only Look Once (YOLO) real-time object detector to generate frequent false-positive detections. To address this issue, this study proposes a serial cascaded detection method that integrates an improved YOLO detector with machine-learning-based secondary verification. In the YOLO branch, channel-prior convolutional attention (CPCA) and learnable weighted multiscale feature fusion are introduced to enhance target representation under complex background conditions and generate candidate bounding boxes. In the machine-learning branch, handcrafted features describing texture, color, shape, edges, morphology, and frequency-domain characteristics are extracted from the candidate regions. A verifier selected through multi-model comparison and ensemble evaluation is then employed to confirm the YOLO-generated candidates. For parameter optimization, the operating point of the standalone YOLO detector with the highest F1-score is first selected as the baseline. A two-dimensional grid search is subsequently performed over 95 threshold combinations consisting of five YOLO candidate-confidence thresholds and nineteen machine-learning confidence thresholds. The optimal configuration is determined using a weighted improvement score defined according to the relative changes in precision, recall, and the F1-score with respect to the baseline. The best overall performance is achieved when the YOLO and machine-learning confidence thresholds are set to 0.25 and 0.75, respectively. Compared with the standalone YOLO detector, the proposed cascaded model improves accuracy from 93.3% to 94.1%, precision from 90.1% to 92.6%, and the F1-score from 93.4% to 94.0%, while recall decreases slightly from 97.0% to 95.5%. These results demonstrate that interpretable local features can effectively filter out false-positive YOLO candidates, thereby suppressing false alarms and improving overall discrimination performance at the cost of only a limited reduction in recall. The developed system has been deployed on rockfall-prone sections of highways G210 and G108, providing technical support for real-time road rockfall monitoring and early warning.