Rockfall detection using lidar point clouds: identification of geometric distortions during acquisition and proposed processing to enable a low detection threshold

The C2R-IA project (www.c2r-ia.fr) is aimed to better account for the influence of weather conditions on the level of rockfall hazards and to anticipate temporary increases in hazard levels during storms and other specific weather conditions, in order to implement risk mitigation systems (access restrictions, monitoring, mobilization of emergency kits, predictive maintenance). To achieve this, a database of rockfall events is built to train AI predictive models of rockfalls based on weather conditions. One of the monitoring technologies used is a terrestrial laser scanner with a RIEGL VZ-2000i long range 3D laser scanning system. Lidar point clouds are thus used to provide at several time intervals the 3D surface of the study site: the Saint-Eynard cliff, located northeast of Grenoble in the french Alps. From the lidar point cloud series, the goal is to compare the clouds to detect changes and identify rockfall events (Manceau et al, EGU 2025, oral presentaion). For a large and rich database, it is important to achieve very precise alignement between lidar point clouds to detect the smallest possible changes in our point clouds series (small rockfall volumes).

In this context, a basic ICP (Iterative Closest Point) alignement reveals artefacts that need to be treated in a special way to achieve high-precision alignement. Geometric distortions are thus observed within the  point clouds in the form of vertical strips. This phenomenon occurs at two scales:

- Low frequency: observations of decimetric to multi-decimetric jumps with strip widths ranging from 10 to 100 meters during acquisitions from a tripod, a flexible support.
- High frequency: observations of centimetric jumps with narrower strip widths (ranging from one to several meters) during acquisitions from a rigid base (reinforced concrete post).

Several hypotheses are put forward and tested to explain the existence of these strips: machine-related mechanical issues, independent or dependent on time, interaction between the ground, support, and machine, changes in atmospheric conditions during the acquisition period (lasting 40 minutes), the geometry of the cliff and its local orientation relative to the lidar's line of sight.

A processing method is proposed to overcome these geometric distortions during acquisition and maintain a low detection threshold when comparing two point clouds: this involves a new strip-based alignment of the two clouds before change detection. The first step is the extraction of strips from the compared cloud, then an independent alignment of each strip to the reference cloud is performed using the ICP method. Finally, the aligned strips are merged to form the new compared cloud : we reach a detection threshold of less than 10 cm (i.e. 10^-4 times the measurement distance) whereas 40 cm has been previously used on the same site in the literature.