Retraining-Free Camera Localization in Indoor Point Clouds Using Edges and Normals

Abstract

Estimating a 6-DoF camera pose in a prebuilt map is essential for spatial perception in robotics and autonomous driving. While structure-from-motion (SfM) maps provide visual context, they lack geometric precision and scale. In contrast, point cloud maps from LiDAR or depth cameras offer accurate geometry for metric-level perception. This article presents a novel camera localization method that enhances pose estimation accuracy in indoor point cloud maps that do not contain visual information. By leveraging low-level features such as edges and normals from both the query image and the point cloud, the proposed approach operates across diverse environments without scene-specific training. Specifically, we introduce an edge alignment cost that measures the edge point reprojection errors and a normal distribution cost using normalized information distance (NID) to quantify normal vector similarity. An adaptive weighting scheme integrates these costs in an optimization framework. Experiments on simulated and real-world datasets demonstrate that our method outperforms state-of-the-art feature-matching algorithms in localization accuracy.