Rockfall object detection on mountain roads based on transfer learning and multi-strategy improvements

To address the challenges of scarce real-world rockfall samples, highly variable target scales, and complex background interference in mountain road rockfall detection, existing models suffer from weak generalization ability and high missed detection rates under small-sample conditions. Consequently, they fail to meet intelligent monitoring requirements. This study aims to construct a rockfall object detection model capable of adapting to small samples, multiple scales, and complex environments, thereby improving detection accuracy and providing technical support for intelligent rockfall monitoring and early warning on mountain roads.

This study proposed a YOLOv8 rockfall detection model integrating transfer learning and multi-strategy improvements. First, a dataset containing over 3 000 rockfall images was constructed, covering various rock lithologies, different lighting conditions, and varying backgrounds. Data sources included on-site surveillance camera captures, mobile phone photography, and data augmentation operations, thereby expanding data diversity from a small-sample perspective. In terms of model design, transfer learning based on a large-scale dataset was introduced to alleviate insufficient feature learning caused by limited samples. Additionally, the coordinate attention mechanism (CA) was embedded into the feature network to enhance the model's response to the spatial positions of rockfalls while suppressing complex background interference. Meanwhile, a weighted bidirectional feature pyramid network (BiFPN) was adopted to replace the original feature fusion path, improving multi-scale feature flow and enhancing the perception of rockfalls at different scales. Finally, the EIoU loss function was employed to optimize bounding box regression, simultaneously considering center point distance, aspect ratio, and overlap area to improve localization accuracy.

Experimental results showed that compared with the baseline YOLOv8 model, the precision of the proposed model increased by 17.1%, recall by 24.7%, and mAP50 by 17.4%, with significant improvements in all three indicators. In comparison with other lightweight models, the proposed model achieved an mAP50 of 74.6%, significantly outperforming YOLOv5n (51.3%), YOLOv6n (56.7%), and YOLOv8n (57.2%). Under small-sample conditions, the missed detection rate and false alarm rate were substantially reduced, and a good balance between accuracy and efficiency was maintained. Validation on the test set demonstrated that the model exhibited good generalization ability and robustness across various complex backgrounds and different rock lithology scenarios.

Through transfer learning and multi-strategy improvements, the proposed YOLOv8 model effectively overcomes the problems of sample scarcity, scale variability, and background complexity in mountain road rockfall detection, achieving outstanding detection accuracy and robustness. It provides a feasible technical solution for the development of intelligent rockfall monitoring and early warning systems for mountain roads.