Motion From Structure: Robust Multi-Image, Multi-Object Pose Estimation

We propose a robust algorithm for tracking camera motion in unconstrained urban environments. We achieve this by using both an attached laser scanner, which gives sparse 3D structure for each frame, and image-based correspondences from a video camera. The combination of a set of 3D points and their associated projections in each frame can be used to find the motion between frames. We propose a new robust multi-object, multi-image pose estimation algorithm and show how to use this algorithm to accurately track the motion of a moving vehicle with an attached camera and 2D laser scanners. Our algorithm is based on a RANSAC formulation and is thus robust to incorrect 3D data and image occlusions, as evinced by tests using synthetic and real data.