Automated classification of piping components from 3D LiDAR point clouds using SE-PseudoGrid

Piping components constitute a crucial part of commercial buildings and industrial plants, and 3D modeling of such components from LiDAR point clouds has attracted wide attention. One of the essential challenges of 3D modeling from point clouds is in classifying object shapes. Unfortunately, accurate and robust classification of complex piping components is difficult and professional interpretation generally needs experts to hand-craft descriptors, which is inefficient. Recently, deep learning (DL) based methods for point cloud shape classification have been successfully applied in the construction field, and several attempts have been made on the classification of piping components. However, conventional DL based methods provided limited support to exploit local context and leverage deep residual learning for recognizing features, leading to inefficient representation learning on point clouds and inferior performance. To overcome these challenges, this paper presents a new DL framework, with the capability of establishing squeeze-excite (SE) mechanism in local aggregation operators and exploiting deep residual learning for point cloud learning, to help classify complex piping components more efficiently and robustly. First, a sole public piping dataset named Pipework is selected, canonicalized and benchmarked using six representative DL-based methods. Second, a new local aggregation operator (LAO) named SE-LAO is developed to extract discriminant features from the point cloud by integrating with the SE mechanism. Finally, a new network termed SE-PseudoGrid is proposed by replacing the LAOs of baseline PseudoGrid with the new SE-LAO. To validate the proposed network, comprehensive experiments are conducted on the Pipework dataset to compare its effectiveness and efficiency with the baseline. Our SE-PseudoGrid outperforms the baseline by a noticeable margin, achieving an overall accuracy (OA) of 96.3% and average class accuracy (avgAcc) of 97.5%. This study can contribute to automated object recognition of piping components from point clouds and facilitate the creation of as-built BIMs.