Self-Supervised Learning Method for Semantic Segmentation of LiDAR Point Clouds

Semantic segmentation has shown a significant success for achieving comprehensive scene understanding in real-time perception and urban modeling. Over the recent years, there have been significant advancements in semantic segmentation for LiDAR point clouds, largely the adopting of deep learning techniques. There are the related works of 3D semantic segmentation, including neural network models to process converted voxels, points, and graphs. However, point-based methods are not taken into account local structure feature, resulting in a lack of fine-grained features and limited generalization. Additionally, these models do not take full advantage of the high-level geometric correlations among local neighbors, resulting in low semantic segmentation accuracy. The use of voxel-based methods for balancing precision and computational efficiency is a useful technique. Still, voxel-based representation of point clouds is inefficient and tends to ignore fine details. Little research has investigated using graph-based methods for LiDAR point cloud semantic segmentation. Our work demonstrates the feasibility of using graph representation for highly accurate semantic segmentation in a point cloud. Other problems in the existing methods are high computational and memory requirements. Self-supervised learning on large unlabeled datasets is one way to reduce the number of manual annotations needed. In this thesis, we explore that leverage the combination self-supervised contrastive learning and graph-based method to overcome the semantic segmentation challenges of large-scale point clouds. An experimental qualitative and quantitative analysis of our method shows that the proposed approach can beat previous approaches on S3DIS and SemanticKITTI datasets for the task of LiDAR point cloud semantic segmentation.

Name thesis:degree_name

Master of Science (MSc)

Discipline thesis:degree_discipline

Engineering – Geomatics

Grantor dc:publisher.institution

Graduate Studies

Year dc:date.issued

2024

Author dc:creator

• Mutlu Kipirti, Fatma

Advisor dc:contributor.advisor

• Wang, Ruisheng

Committee members dc:contributor.committeemember

• Wang, Ruisheng
• Hassan, Quazi Khalid
• Wang, Xin
• Yang, Hongzhou

• LiDAR point clouds
• self-supervised learning
• semantic segmentation

Statement dc:rights

• University of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.

Language dc:language.iso

en

DOI

https://doi.org/10.11575/PRISM/43341

OAI identifier oai:identifier

oai:ucalgary.scholaris.ca:1880/118499