Intelligent classification of point clouds for indoor components based on dimensionality reduction

Abstract

With the wide application of LiDAR, RGBD cameras and other sensors in computer vision, intelligent robotics, indoor positioning and navigation, the processing of point clouds of indoor scene components has been a difficult problem in these fields. Due to the disorder, sparsity, and limited information of point clouds, it is a challenge to consume point cloud directly. This paper proposes an intelligent classification method based on the disordered point clouds of indoor components. First, a deep learning network is employed to extract high-dimensional features. Then the features are divided into different clusters using two algorithms: t-distributed stochastic neighbor embedding (t-SNE) and density-based spatial clustering with applications of noises (DBSCAN). Finally, the classical iterative closest point (ICP) is used to match the laser point clouds with the model point clouds whose semantic labels are known in the model dataset. As a result, the method has a good performance on the classification of indoor point clouds, and the accuracy of classification is 98.6%, which can realize the intelligent classification of indoor components point clouds.