Robust Probabilistic Occupancy Grid Estimation from Positive and Negative Distance Fields

We present an approach for estimating occupancy grids with an emphasis on robotics applications, where collision avoidance and robustness to severe noise are of more importance than high resolution. We build upon probabilistic techniques, typically used in robotics, and techniques based on signed distance fields, typically used in computer vision, to obtain an approach that is robust and also allows probabilistic reasoning on free and occupied space. The uniqueness of our method lies in the use of separate accumulators for positive and negative evidence for the occupancy of each voxel. This enables our representation to capture the uncertainty due to potential conflicts among the measurements instead of allowing contradictory evidence to cancel each other out. We show occupancy grids computed from multi-view stereo inputs on precisely and imprecisely calibrated image sequences. The ground truth that is available with the former dataset allows quantitative evaluation of the performance of our algorithm.