Assessing the impact of 3D point neighborhood size selection on unsupervised spall classification with 3D bridge point clouds

Abstract

The performance of point cloud-based defect classification algorithms depends on the quality of the computed geometric features, which in turn is strongly affected by the selection of the neighborhood size of local 3D points. Many existing algorithms select a single scene-specific value as the neighborhood size parameter and apply it uniformly to all 3D points while computing features. By doing so, some defect features end up being smoothened out at locations where the selected neighborhood size is larger than the optimal choice for those points. Conversely, geometric features might not always capture the local 3D structure if they are calculated based on less than optimal number of neighbors. This paper investigates and assesses the relationship between neighborhood size selection and the performance of point cloud-based unsupervised spall classification algorithms in a quantitative manner. Among the presented neighborhood selection approaches, an entropy-based approach that incorporates a tailored optimal neighborhood size for every point in a point cloud, resulted in a significant improvement over the performance of current state-of-the-art approaches. The performed quantitative study also demonstrated the robustness of this approach to variables, such as subsampling percentage, maximum neighborhood size and different noise levels. The implemented research testbed is comprised of statistically significant number of spall defect datasets generated from five different bridges. The outcome of this research is expected to improve the reliability of point cloud-based condition assessment for concrete bridges.