Spatiotemporal Rockfall Detection Using Point-Based Neural Networks

Rockfall poses a significant threat to human life and infrastructure in mountainous regions, necessitating effective detection and mitigation strategies. Sensor technologies, such as Terrestrial Laser Scanners, are being widely utilized to periodically scan mountain terrains and cliffs acquiring 3D point clouds. Current research on intelligent rockfall detection utilizes pre-computed features and machine-learning models, clearly lacking hidden geometric properties inherent in 3D point clouds. Also, recent point-based deep learning approaches that focus on geometric feature extraction and end-to-end learning, mainly study datasets with balanced labeled observations, not addressing the rockfall class imbalance in real-world cases. Our approach builds upon advancements in point-based neural networks and integrates spatiotemporal information to enhance accuracy and efficiency in detecting rockfall candidates in cases where rockfall observations are limited. Addressing the class imbalance issue inherent in rockfall detection, we present results in real-world 3D scans from a cliff in Spain showcasing the effectiveness of our method in accurately identifying rockfall events.