Midterm colloquium Christian Vorage

Point Cloud Compression for Automotive LiDAR using Tensor Decomposition Methods

Supervisors:  Dr.ir. K. Batelier & Dr. J.F.P Kooij

Abstract: In recent years, the automotive industry has made numerous advancements towards autonomous driving. These advancements can partially be attributed to developments in sensor technologies such as Light Detection and Ranging (LiDAR), which enable a vehicle to perceive and navigate its surroundings. Unfortunately, the sheer volume of these LiDAR scans (~100.000 points per scan) poses significant challenges for self-driving vehicle systems, both in training and in operation. Hence, a key element which aids the transition to a self-driving future, is the development of algorithms capable of efficiently compressing LiDAR scans with negligible reconstruction loss.

This work focuses on an alternative approach for compressing automotive LiDAR point clouds, based on concepts from multilinear algebra. This approach is known as tensor decomposition, and is often employed in analysis of multi-dimensional data such as time-series, images or videos. Three different tensor decomposition methods are considered: Canonical Polyadic Decomposition, Tucker Decomposition and Tensor Train Decomposition.