Student Name

Zeerak A. KakarFollow

Date of Submission

Fall 2025

Supervisor

Dr. Faisal Iradat, Assistant Professor, Department of Computer Science, School of Mathematics and Computer Science (SMCS)

Co-Supervisor

Dr. Pengpeng Hu, Associate Professor, Department of Materials Faculty of Science and Engineering, University of Manchester

Committee Member 1

Dr. Taslim Murad, Examiner – I, Institute of Business Administration, IBA Karachi

Committee Member 2

Dr. Muhammad Umer Farooq, Examiner – II, Department of Computer Science & IT, NED University

Committee Member 3

Dr. Muhammad Atif Tahir, Professor and Program Coordinator, Graduate & Postgraduate Programs (CS)

Degree

Master of Science in Data Science

Department

Department of Computer Science

Faculty/ School

School of Mathematics and Computer Science (SMCS)

Keywords

3D Point Clouds, Surface Area Estimation, Point Cloud Registration, 3D  Reconstruction, Synthetic Dataset, Deep Learning.

Abstract

Point cloud registration remains a challenging problem in computer vision and remote sensing due to sparsity, noise, and the lack of reliable global geometric cues. Existing methods largely depend on local correspondences and iterative optimization, which are sensitive to initialization and computationally expensive. Additionally, most real-world point cloud datasets do not provide ground-truth surface area information, limiting the use of surface-based geometric constraints. This research proposes a Surface Area–based Reduction Module (SARM) that introduces surface area as a global geometric feature to support point cloud registration. A large-scale synthetic terrain dataset containing over 200,000-point cloud samples was generated, representing diverse natural landscapes, with accurate surface area ground truth obtained through mesh-based triangulation. Several deep learning models were evaluated for surface area estimation, including voxel-based 3D CNNs, PointCNN, DGCNN, and PointNet++. Experimental results show that PointNet++ achieves the best performance, benefiting from hierarchical feature learning and multi-scale neighborhood aggregation. The proposed framework attains an RMSE of 4.76%, demonstrating robust and accurate surface area estimation across varying terrain types. The results confirm the effectiveness of SARM and highlight its potential for improving point cloud registration and related applications in terrain analysis and remote sensing.

Document Type

Restricted Access

Submission Type

Thesis

Share

COinS