Prototype development of on-board vision processing micro air vehicle for wide-range monocular SLAM

This study was undertaken to create an indoor environmental map using a monocular camera based on structure from motion (SfM). Numerous methods on simultaneous localization and mapping (SLAM) have been proposed for 3D map construction used for various types of mobile robots. We used MonoSLAM proposed by Davidson et al., which is a fundamental method using extended Kalman filters (EKFs) for feature tracking and updating with camera position estimation. For this study, we developed our original micro air vehicle (MAV) prototype to obtain time-series images in a wide-range of indoor environments. The MAV prototype has four monocular cameras to actualize rapid omnidirectional sensing and has computer systems for online vision processing. We applied the method to a benchmark dataset and our original dataset, which was obtained using a single monocular camera. Using the benchmark dataset, we confirmed the appearance tendencies and their meaning of three color patches: red, blue, and yellow patches that respectively correspond to correct tracking, failure tracking, and temporary waiting local features. The experimentally obtained results obtained using our dataset reveal detection of numerous yellow patches, a few blue patches, and a few red patches. Moreover, we obtained 3D trajectories of a camera on the MAV as 3D maps.