Accurate Pose Refinement of Detected Vehicles Using LiDAR Point-to-Surfel ICP and Vehicle Shape Priors

This letter proposes an accurate pose refinement system that integrates point cloud registration and vehicle prior shapes to improve LiDAR-based vehicle pose estimation. For safe autonomous driving, centimeter-level accuracy is essential for estimating the poses of nearby vehicles. Although existing vehicle detection algorithms have achieved high performance with the advancement of deep learning, they still struggle to capture the fine-grained movements of surrounding vehicles at the centimeter level. To address this challenge, we propose an algorithm that refines a vehicle's pose estimate by registering its LiDAR point cloud with a prior vehicle shape. First, we obtain a 3D bounding box from an existing object detection module using the LiDAR point cloud data. Next, we perform image classification within that bounding box to identify the specific vehicle model. We then retrieve the corresponding 3D surfel model from a car shape database and perform a point-to-surfel ICP (Iterative Closest Point) alignment with the point cloud of the actual vehicle. Through the optimization process, we acquire a more precise pose estimate than the existing object detection module. Our approach can improve the performance of various tasks that utilize object detection information. In this letter, we demonstrate its effectiveness by integrating the refined pose estimates into a vehicle detection and tracking (VDT) experiment. The experimental results show that our proposed pose refinement module can also enhance tracking performance, further validating the advantages of our method.