Aplicação de deep learning para segmentação e detecção de detritos em nuvens de pontos

Abstract

The automatic detection of cracks in LiDAR point clouds of concrete structures faces a critical practical challenge: the scarcity of labeled data, since manually annotating each point as crack or normal surface is infeasible at scale. This work proposes and evaluates eight machine learning methods for this problem, covering the full supervision spectrum supervised, semi-supervised, and unsupervised, all trained in global mode and evaluated under the Leave-One-Cloud-Out (LOCO) protocol on a dataset of 64 point clouds from the Igreja Nossa Senhora do Rosário (Aracati, Brazil), acquired with the Leica BLK 360 scanner. The results demonstrate that global training, in which data from all training clouds are concatenated into a single set, is the most determining factor for effective detection, outweighing differences in architectural complexity. The GMMScalar, a multilayer perceptron trained on 18-dimensional scalar and geometric features, achieved the best mean AUROC (0.976±0.023), surpassing PTv3-Binary (0.966) and ScalarGAT (0.929). Migrating from per-cloud to global training improved the AUROC of the GMM methods by up to 36 percentage points. Diagnostic analysis identified two discriminability regimes in the scalar field, clouds with clear separation (gap) and clouds with overlap (overlap), which explain the variability of the results and quantify the cost of absent supervision (ΔAUROC = 0.35).