Real-time Simultaneous Localization and Mapping using Omnidirectional Vision

This paper presents an Extended Kalman filter (EKF) based real-time simultaneous localization and mapping (SLAM) method for a robot with an omnidirectional camera. In the EKF part, we employ a sequential process that utilizes features in turn for updating the robot pose estimate. This approach enables us to efficiently reduce the computational complexity and predict feature search regions. Thus, it contributes to run the Kalman recursion in real time. In the feature management part, we deal with the initialization and measurement of features. To initialize the covariance of the feature as well as its 3D position, we exploit the uncertainties of robot pose and image noise at the same time. For robust measurement, features are reordered according to their qualities of cornerness and covariance. Experiments for the real-time localization and map building with an omnidirectional camera are demonstrated in an indoor environment. The computation time is about 45Hz when using around 30 features. Robot pose estimates are compared with the ground truth in terms of accuracy.