Tools
Tools
Plunkett, Hugh Christopher (2025) Lidar based 3D Detection of Vehicles and Obstacles in Autonomous Vehicles Using Deep Learning Techniques: Analysis and Comparison. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (2MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (850kB) | Preview |
Abstract
Autonomous Vehicles (AVs) face significant challenges in terms of accurate and real-time 3D object detection, which is critical for autonomous driving. This study aims to explore and compare several prominent state-of-the-art deep learning (DL) approaches on the KITTI dataset. Using a multifaceted evaluation approach, we conducted a standardized analysis of the most recent LiDAR-based 3D detectors on the dataset. Our work uniquely integrates hybrid models into the comparison and addresses real-world deployment challenges by evaluating computational trade-offs between speed and accuracy. A thorough comparison of LiDAR-based 3D detectors is needed to assess their accuracy and speed for real-world AV deployment. Among the eight models evaluated, PointRCNN and VoxelRCNN achieved the highest accuracy, followed closely by Complex Yolo v4 and PV-RCNN, indicating strong performance in 3D detection tasks. The results highlight the better performance of voxel-based and point-based deep learning models over traditional approaches in both accuracy (71 and 68 mAP compared to AVOD’s 47 mAP) and efficiency (VoxelNet runs at 46 FPS, 3x faster than AVOD while being more accurate). All models have been tested in a standardised environment with the hope to provide valuable insights for selecting models, advancing research toward practical, high-performance perception systems in AVs.
Actions (login required)
 |
View Item |