Scholarlink

Autonomous vehicles (AV) are a gamechanger in the technology of today. Capable of driving without constant human attention, AVs rely on various sensors to constantly scan the environment to ensure smooth navigation. The LiDAR sensor (Light Detection and Ranging) is the key sensor to detect obstacles on the road. While it is robust against illumination unlike cameras, the LiDAR sensor produces raw point clouds. Deep learning allowed them to be analyzed for vehicle perception. Due to that, 3D object detectors for LiDAR point clouds are developed to help improve vehicle perception. Pretraining is one of the methods used to reduce development time as it allows existing data to be transferred from one source to another domain. In this research, two pretraining techniques, self-supervised learning and transfer learning are applied onto an existing 3D object detector using the famous KITTI dataset as the source domain. Based on results obtained from both techniques, the transfer learning model performed better than the self-supervised learning model in terms of mAP (mean Average Precision) as the KITTI dataset is already well annotated. While for actual deployment, both transfer learning model and self-supervised learning model are feasible in detecting additional objects in the LiDAR pointclouds.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparison Between Transfer Learning and Self-supervised Learning Techniques for Van and Truck Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ericsson Yong Wing Mei,
  • Muhammad Aizzat Zakaria,
  • Mohamad Heerwan Peeie,
  • M. Izhar Ishak

摘要

Autonomous vehicles (AV) are a gamechanger in the technology of today. Capable of driving without constant human attention, AVs rely on various sensors to constantly scan the environment to ensure smooth navigation. The LiDAR sensor (Light Detection and Ranging) is the key sensor to detect obstacles on the road. While it is robust against illumination unlike cameras, the LiDAR sensor produces raw point clouds. Deep learning allowed them to be analyzed for vehicle perception. Due to that, 3D object detectors for LiDAR point clouds are developed to help improve vehicle perception. Pretraining is one of the methods used to reduce development time as it allows existing data to be transferred from one source to another domain. In this research, two pretraining techniques, self-supervised learning and transfer learning are applied onto an existing 3D object detector using the famous KITTI dataset as the source domain. Based on results obtained from both techniques, the transfer learning model performed better than the self-supervised learning model in terms of mAP (mean Average Precision) as the KITTI dataset is already well annotated. While for actual deployment, both transfer learning model and self-supervised learning model are feasible in detecting additional objects in the LiDAR pointclouds.