Rockfall-YOLO Deep Learning Model for Rockfall Motion Object Detection

作者：Guoshao Su, Weijie Huang, Chongjin Li, Xiaojie Huang · 发表于：Journal of Computing in Civil Engineering · 年份：2025 · DOI：10.1061/jccee5.cpeng-6268 · 被引用次数：2 · 研究领域：Landslides and related hazards、Anomaly Detection Techniques and Applications、Video Surveillance and Tracking Methods

A Rockfall-You Only Look Once (YOLO) deep learning model is proposed in this study for rockfall detection to overcome the problems of current rockfall object detection models, such as the low detection accuracy of motion targets with blurred contours, the inability to distinguish between motion and static rock masses, susceptibility to smoke interference, difficulty in detecting rockfalls at extremely small scales, and failure to reasonably estimate the volume of rockfalls. The proposed model is based on the YOLOv5 model. An edge enhancement module based on the Sobel operator is introduced to highlight the edge contour information of the target to be detected. A motion filtering module is introduced to the detection frame generation to enable the model to focus on the detection of moving targets. In the sample training, smoke is introduced as the detection target to improve the confidence of falling rock detection. The feature pyramid networks (FPNs) and pixel aggregation network (PAN) structure is improved through network deepening and lateral information transfer path change, and some of the convolution modules are replaced with a residual component to improve the detection of extreme small-scale targets. In addition, a volume estimation module based on the monocular ranging principle was introduced, a sample dataset was created from test footage of rockfalls on slopes and images of rockfalls collected by the network, and the deep learning model was trained accordingly. The...