시각 관성 주행 거리 측정을 사용한 위치, 자세 추정 및 3D 지도 작성

Abstract

Map building and real-time localization are essential for the autonomous navigation of mobile robots. This paper shows an implementation of mapping and localization algorithm and analysis of the performance of the algorithm. The implementation uses multi-state constraint Kalman filter(MSCKF) for localization and RTAB-Map for mapping. This study uses ceiling-view image and floor-view image, whereas previous applications of MSCKF and RTAB-Map generally use front-view image for localization. The implementation calibrates the camera and inertial measurement unit(IMU). The calibration includes transformations among the cameras and IMU. It also uses Allan variance approach to adjust the covariance of measurement errors. The trajectories by the MSCKF were compared to investigate the effect of viewing direction to the localization performance of VIO and mapping performance of the RTAB-Map. The analysis show that number and characteristics of feature affect the performance. Even if there are large number of featured points recognized, the performance degrades if correspondence of features are not correctly identified. The results of the study can be used for adequate implementation of navigation algorithms in 3-Dimensional space.