Analysis of the impact of TLS point cloud feature sets on the detection of building displacements using machine learning algorithms

Abstract The study addresses the problem of detecting displacements of building walls using terrestrial laser scanning (TLS) data and machine learning methods. Traditional displacement measurement techniques are often time-consuming. They are also limited in capturing the full geometry of monitored objects. In response, this research proposes a methodology based on the analysis of geometric and radiometric features extracted from point clouds. Controlled experiments were conducted with a geodetic rosette equipped with distance-measuring prisms, which were displaced in the XY plane by 4 mm, 9 mm, and 13 mm. Data were recorded with a Leica RTC360 scanner from three stations, yielding nine point clouds. Selected features describing differences between corresponding points in the reference and displaced series were used as input for neural network models. Both binary classification (displacement/non-displacement) and multi-class classification (0, 4, 9, 13 mm displacement) were performed. The results demonstrated high classification accuracy: 99.1% for binary models and 96.0% for multi-class models. Feature ranking revealed that geometric attributes, such as displacement vector length, curvature, and normal vectors, were the most relevant for model training, while color features had minor importance. The study confirmed that scanner position and incidence angle of the laser beam strongly affect classification quality. The developed procedure proved effective in detecting displacements that occur in directions parallel to the plane of building walls. The conclusions drawn from the research constitute a valuable contribution to the theory of monitoring building structures using TLS.