VoxFall: Non-parametric volumetric change detection for rockfalls

Abstract

Surveying methods such as digital photogrammetry and laser scanning have been used to detect surficial changes by comparing two 3D models. This study deals with the manipulation of 3D digital models of rock slopes within the scope of rockfall monitoring which includes the objective of detecting and quantifying rockfall events as discrete blocks. Current change detection methods for rockfalls are based on distance computation between two rock slope models complemented successively by spatial clustering and cluster shape reconstruction routines, and include severe challenges associated with the profound interdependence of parameter tuning between the different steps. To solve these issues, we introduce a new algorithm – VoxFall – that does not rely on distance computation and its objective is to eliminate user subjectivity by launching a new tool for rockfall monitoring that would only be controlled by the quality of the input data. The method treats the two input models as a single scene and applies two steps: 1) fitting an occupancy voxel grid of a resolution defined by the registration error; 2) empty space clustering and volume computation based on voxel adjacency. Comparison with existing methods across both synthetic and real rock slope datasets demonstrates the sensitivity of the distance-based methods and the dependency on the input parameters compared to the results of our method. Application on original data predicts almost perfectly the rockfall volume (0.3 % difference) within an arrangement of recorded rockfall events. We provide evidence of current techniques requiring pre-existing knowledge of rockfall activity to tune them while VoxFall comprises a unified framework that enables direct accurate volume detection and clustering with no user intervention. The algorithm has been implemented in an open-source software package.