Sharing Heterogeneous Spatial Knowledge: Map Fusion Between Asynchronous Monocular Vision and Lidar or Other Prior Inputs

Abstract

To enable low-cost mobile devices and robots equipped with monocular cameras to obtain accurate position information in GPS-denied environments, we propose to use pre-collected lidar or other prior data to rectify imprecise visual simultaneous localization and mapping (SLAM) results. This leads to a novel and nontrivial problem that fuses vision and prior/lidar data acquired at different perspectives and time. In fact, the lidar inputs can be replaced by other prior mapping inputs as long as we can extract vertical planes from these inputs. Hence, they are referred as prior/lidar data in general. We exploit the planar structure extracted from both vision and prior/lidar data and use it as the anchoring information to fuse the heterogeneous maps. We formulate a constrained global bundle adjustment using coplanarity constraints and solve it using a penalty-barrier approach. By error analysis we prove that the coplanarity constraints help reduce the estimation uncertainties. We have implemented the system and tested it with real data. The initial results show that our algorithm significantly reduces the absolute trajectory error of visual SLAM by as much as 68.3%.