A DGCNN-Based Framework for Indoor Object Classification Using Sparse 2D LiDAR Point Clouds

Abstract

Object classification using two-dimensional Light Detection and Ranging (2D LiDAR) remains challenging because the resulting point clouds are sparse, irregular, and unordered. This study investigates a Dynamic Graph Convolutional Neural Network (DGCNN)-based framework for classifying indoor objects from sparse 2D LiDAR scans. The framework integrates polar-to-Cartesian transformation, invalid-point filtering, fixed-size point sampling, centroid-based normalization, dynamic graph construction, and EdgeConv-based feature extraction. A locally collected dataset containing 783 point-cloud samples from four object categories, namely chair, table, person, and sofa, was used for evaluation. The data were acquired using a single-line 2D LiDAR sensor in a controlled indoor environment. The evaluated model achieved an accuracy of 0.98 on the testing dataset. A comparison with a PointNet++ baseline under the same experimental configuration achieved 0.97 accuracy, demonstrating comparable capability between dynamic graph-based and hierarchical pointcloud learning approaches. The results indicate that DGCNN can effectively represent sparse planar LiDAR geometry while remaining limited by ambiguous object structures caused by incomplete vertical information.