Segmentation of pipelines and power cables from point clouds using a deep learning approach

Details

Supervisors

Schaus, Pierre

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

This master's thesis aims to develop a novel deep learning approach for the accurate segmentation of pipelines and power cables from point clouds, addressing the challenges of infrastructure monitoring and enhancing construction processes. The research is conducted in collaboration with Space Time, an innovative company specializing in 4D scanning services. The model utilizes a shifting local view mechanism to accurately track the path of pipelines and extract relevant features over a custom dataset. This dataset was meticulously developed to suit the specific requirements of the pipeline and power cable segmentation task. The integration of PointNet, a renowned module for point cloud segmentation, initially showed suboptimal performance. Consequently, a thorough theoretical analysis was conducted, providing extensive evaluations that clearly demonstrate the effectiveness of the model in accurately segmenting pipelines and power cables. The findings contribute to the field of point cloud processing and showcase the lessons learnt in the application of deep learning techniques for complex pattern recognition tasks.