3D LiDAR-Based 6D Pose Estimation for Object Trajectory Prediction in Autonomous Vehicle Applications Minjae Cho School of Mechanical Engineering Gwangju Institute of Science and Technology

Abstract

An autonomous driving system is a technology that enables a vehicle to navigate itself to a given destination without human intervention. This technology is being ac- tively researched due to its promise to reduce tra!c accidents, improve transportation e!ciency, and save time and energy. In autonomous driving, vision sensors such as cameras and LiDAR detect surrounding objects and predict their trajectories. Camera sensors are currently the most commonly used, but they do not provide direct distance information, which limits their use for safe driving. LiDAR sensors, on the other hand, provide direct distance information through 3D point cloud data, enabling safer driving. To achieve high safety and e!ciency in autonomous driving systems, it is necessary to predict the trajectories of objects. Currently, LiDAR-based 3D object detection is widely used to predict the trajectory of objects. These techniques are mainly catego- rized into voxel-based, point-based, and fusion-based methods. Voxel-based methods convert sparse 3D point cloud data into regular 3D voxels or 2D grids and perform 3D sparse convolutional neural networks (CNNs), which are computationally e!cient but have the limitation of low location accuracy. Point-based methods focus on all raw point cloud data to preserve accurate location information but could be more compu- tationally e!cient because they focus on all points. Fusion-based methods combine the advantages of both, preserving computational e!ciency and location information. Conventional LiDAR-based 3D object detection techniques estimate an object’s class, size, and z-axis rotation, ignoring the x-axis and y-axis rotation information. However, 6D pose estimation is essential to accurately predict the trajectories of all objects, including two-wheelers and pedestrians. In this paper, I propose two meth- ods for estimating a 6D pose of objects. First, I propose constructing a synthetic 6D pose dataset in a simulation environment for training the network. Second, I propose Point-Voxel based-Pose estimation networks (PV-Pose) for 6D pose estimation on point clouds by applying transfer learning based on the PV-RCNN++ model. ©2025 Minjae Cho ALL RIGHTS RESERVED