Deterministic Visual Mapping

Pierre Bénet and Alexis Guinamard

Abstract: This paper presents our development of monocular visual Simultaneous Localization And Mapping (SLAM), and its extensions to inertial aiding and stereo. Following our development of stereo visual odometry RADVO, we were able to apply on our monocular visual SLAM system the same visual feature extraction and matching, and the same outlier removal principles. Hence the feature matching strategy is done with sparse Normalized-Sum of Squared Difference matching and the outlier removal is based on a quadratic loss function. Which makes our system one of the few to be deterministic. Since our visual features are rather simple and are localized at a pixel, we applied successfully several concepts from direct visual methods such as Direct Sparse Odometry (DSO) or Direct Sparse Mapping (DSM). Hence, we model a feature using its inverse distance to the camera and we use about ten active frames and a thousand active features to perform optimization (bundle adjustment) for short term SLAM. Unlike many SLAM systems, our loop closing strategy relies only on prediction rather than place recognition. Like DSM, this is made possible by our short-term SLAM that is precise enough to predict when an old frame will be visible again. We pushed further our development to be able to perform global bundle adjustment of all the features in all the keyframes of a log, optimizing millions of features in a few seconds for post-processing. We then achieve state of the art precision on the EuRoC dataset and TUM-VI dataset in all sensor configuration. In particular our stereo-inertial SLAM achieves an average accuracy of 2.9 cm on the EuRoC drone in real time. An accuracy that drops to 2.6 cm after a few seconds of post processing using block-sparse conjugate gradient.
Published in: Proceedings of the 34th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2021)
September 20 - 24, 2021
Union Station Hotel
St. Louis, Missouri
Pages: 3173 - 3188
Cite this article: Bénet, Pierre, Guinamard, Alexis, "Deterministic Visual Mapping," Proceedings of the 34th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2021), St. Louis, Missouri, September 2021, pp. 3173-3188. https://doi.org/10.33012/2021.18084
Full Paper: ION Members/Non-Members: 1 Download Credit
Sign In