Accurate Single Image Multi-modal Camera Pose Estimation

A well known problem in photogrammetry and computer vision is the precise and robust determination of camera poses with respect to a given 3D model. In this work we propose a novel multi-modal method for single image camera pose estimation with respect to 3D models with intensity information (e.g., LiDAR data with reflectance information). We utilize a direct point based rendering approach to generate synthetic 2D views from 3D datasets in order to bridge the dimensionality gap. The proposed method then establishes 2D/2D point and local region correspondences based on a novel self-similarity distance measure. Correct correspondences are robustly identified by searching for small regions with a similar geometric relationship of local self-similarities using a Generalized Hough Transform. After backprojection of the generated features into 3D a standard Perspective-n-Points problem is solved to yield an initial camera pose. The pose is then accurately refined using an intensity based 2D/3D registration approach. An evaluation on Vis/IR 2D and airborne and terrestrial 3D datasets shows that the proposed method is applicable to a wide range of different sensor types. In addition, the approach outperforms standard global multi-modal 2D/3D registration approaches based on Mutual Information with respect to robustness and speed. Potential applications are widespread and include for instance multispectral texturing of 3D models, SLAM applications, sensor data fusion and multi-spectral camera calibration and super-resolution applications.