3D point cloud management & compression

• Main Author: Yu, Zhigang
• Other Authors: Lin Weisi
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Data
• Online Access: https://hdl.handle.net/10356/157242
• Institution: Nanyang Technological University

This project aims to explore various data representations for Point Cloud, and algorithms for sparse Point Cloud. Referring to OctSqueeze[1] algorithm, an octree representation is used for Point Cloud which is able to retain required information as much as possible. Then the entropy model is applied to further compress Point Cloud data octree leaf node to the bitstream. The probability distribution for the entropy model is generated by a deep learning model which takes context information available from the octree itself during training, encoding, and decoding. In the ideal algorithm, the main optimization is that it concerns three more context information in the deep learning model compared to OctSqueeze- the tree node occupancy condition at the same level, the intensity information, and the context information in the neighbor frame. Using OctSqueeze as the baseline, the project is trying to optimize the Deep Entropy Model which uses Octree, deep learning, and entropy model to compress the sparse Point Cloud sequential data from the KITTI dataset and try to improve the metrics for accuracy & compression rate compared to the state-of-art algorithm in the same fields.