کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Multi-Scale Geometric Feature Learning from 3D Point Clouds for Automated Defect Classification

نویسندگان :

Sajjad Gharibi¹ Amin Kavousi² Alireza Hadi³

1- دانشگاه تهران 2- دانشگاه تهران 3- دانشگاه تهران

کلمات کلیدی :

Defect Classification،Point Cloud،Multi-Scale Features،Feature Engineering،Machine Learning

چکیده :

This paper presents a classification-centric pipeline for per-point defect detection in sewer pipeline point-clouds, augmented with geometric profiling and visualization for interpretability. Each point is encoded with a 111-dimensional multi-scale geometric descriptor that combines intrinsic coordinates, eigen-based shape measures, multi-scale curvature and density statistics, and multi-scale Fast Point Feature Histograms (FPFH). We evaluate eight classical classifiers on a mixed dataset of 500 real and 800 synthetic scans (≈1.32M labeled defect points) and apply systematic hyperparameter optimization (RandomizedSearchCV on a 100k subsample) to the top candidates. The optimized Random Forest model (500 trees, full depth, min_samples_split=2) achieves Accuracy = 97.22% and Macro F1 = 97.62% on the held-out test set. Beyond classification accuracy, we perform explicit Geometric Profiling and Visualization — including per-class geometric statistics, heatmaps and sectional summaries — to characterize discriminative scales and validate feature relevance prior to modeling. We also profile computational behavior across the pipeline (neighbor search, feature extraction, descriptor assembly, training, inference) on our hardware setup, identifying nearest-neighbor search and feature extraction. The result is a high-accuracy, interpretable, and computationally practical defect-classification solution targeted at resource-constrained inspection platforms