Date of Award
Electrical and Computer Engineering
ADAS, Autonomous, LiDAR, Point Cloud, sensor fusion, vehicle
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.
In the race for autonomous vehicles and advanced driver assistance systems (ADAS), the automotive industry has energetically pursued research in the area of sensor suites to achieve such technological feats. Commonly used autonomous and ADAS sensor suites include multiples of cameras, radio detection and ranging (RADAR), light detection and ranging (LiDAR), and ultrasonic sensors. Great interest has been generated in the use of LiDAR sensors and the value added in an automotive application. LiDAR sensors can be used to detect and track vehicles, pedestrians, cyclists, and surrounding objects. A LiDAR sensor operates by emitting light amplification by stimulated emission of radiation (LASER) beams and receiving the reflected LASER beam to acquire relevant distance information. LiDAR reflections are organized in a three-dimensional environment known as a point cloud. A major challenge in modern autonomous automotive research is to be able to process the dimensional environmental data in real time. The LiDAR sensor used in this research is the Velodyne HDL 32E, which provides nearly 700,000 data points per second. The large amount of data produced by a LiDAR sensor must be processed in a highly efficient way to be effective. This thesis provides an algorithm to process the LiDAR data from the sensors user datagram protocol (UDP) packet to output geometric shapes that can be further analyzed in a sensor suite or utilized for Bayesian tracking of objects. The algorithm can be divided into three stages: Stage One - UDP packet extraction; Stage Two - data clustering; and Stage Three - shape extraction. Stage One organizes the LiDAR data from a negative to a positive vertical angle during packet extraction so that subsequent steps can fully exploit the programming efficiencies. Stage Two utilizes an adaptive breakpoint detector (ABD) for clustering objects based on a Euclidean distance threshold in the point cloud. Stage Three classifies each cluster into a shape that is either a point, line, L-shape, or a polygon using principal component analysis and shape fitting algorithms that have been modified to take advantage of the pre-organized data from Stage One. The proposed algorithm was written in the C language and the runtime was tested on a two Windows equipped machines where the algorithm completed the processing, on average, sparing 30% of the time between UDP data packets sent from the HDL32E. In comparison to related research, this algorithm performed over seven hundred and thirty-seven times faster.
Bondy, Bradford Scott, "3D LiDAR Point Cloud Processing Algorithms" (2020). Electronic Theses and Dissertations. 8295.
Available for download on Tuesday, May 25, 2021