Bayesian network for online global pose estimation

The Problem: Many robot navigation tasks require accurate sensing of the robot’s location. This is a particularly difficult problem when the environment is uninstrumented or has never been explored before. This paper addresses the problem of computing an accurate estimate of the robot’s position given a sequence of scans from an optical sensor such as a monocular camera, a stereo camera, or a laser range scanner. No prior knowledge or instrumentation of the environment is required. Our algorithm computes a globally consistent trajectory, so that when the target returns to an already visited position, its estimated pose is consistent with the pose estimate produced on the earlier visit. To compute the trajectory, we build a non-causal filter which continually updates the past poses as new scans become available, maintaining an optimal trajectory at each time step. Our algorithm does not need to build a 3D model of the world, since it tracks with respect to past frames, and updates the trajectory in time proportional to the number of frames seen so far.