Robust Model-Based Clustering Techniques for Non-Uniform LiDAR Point Clouds via Range Image Transformations

Abstract

Abstract Autonomous driving is a rapidly advancing field that is reshaping the future of transportation by promising improved traffic flow, reduced energy consumption, safer roads, and the elimination of human intervention, thereby saving time. An essential part of autonomous driving is a strong perception system such cameras and radar, which are key for safety features like lane keeping, adaptive cruise control, and emergency braking. However, LiDAR technology stands out being a central part of the perception system due to its ability to provide high-quality 3D spatial information even during poor lighting conditions. A specific type of modern LiDAR sensor (e.g., Luminar Iris) introduces non-uniformity in point cloud distribution by offering adaptive resolution. This feature provides more details for the area of interest, while reducing unnecessary data in less relevant areas such as the sky. However, this non-uniformity of the LiDAR data introduces challenges for classical clustering algorithms, which work under the assumption of a uniformly distributed LiDAR point cloud (i.e., fixed sensor angular resolution). Consequently, these algorithms will struggle with over and under-segmentation problems when the distance between objects varies. In this thesis, two robust model-based clustering algorithms are developed and designed to cluster non-uniform LiDAR point clouds. The methods build on classical breakpoint detection and range image-based clustering approaches to handle varying point densities while still being computationally efficient. Furthermore, two ground removal approaches are developed and evaluated along with a suggestion of two evaluation methods, visual and quantitative evaluation on a dataset of Luminar point clouds. The results demonstrate better clustering performance of the two developed algorithms compared to a baseline algorithm, thus resulting in fewer over-segmentation and under-segmentation errors. Moreover, a thorough analysis of computational is provided. In summary, this research resulted in a stable and reliable clustering performance across diverse scenarios, which helps to realize safer and more predictable autonomous driving without relying on large-scale labeled data for training, as in deep learning models.