Image-database SLAM of an Indoor Robot using RGB-D camera

Conventional VisualSLAM methods use cameras on a vehicle to take environment images during travel while repeatedly observable visual features from these images are tracked so that the camera positions and the feature coordinates can be estimated using nonlinear state observers. Recently, a new visual navigation method using image database has been proposed. In the method, images taken at different locations in a particular environment form a database. Real-time images taken during navigation are compared with database images to determine the relative displacement between the real-time image position and near-by database image positions. If images in the database have correct positions, real-time image position can be determined exactly. In the SLAM approach, new images can be constantly added to the database to increase environment coverage and precise database image positions are not required as they are estimated together with the real-time image/vehicle position using nonlinear state observes. In this paper the implementation and experiment result using RGB-D camera (Microsoft Kinect) and Image-database SLAM method for an omni-wheel robot is presented. While navigating in an unknown environment, RGB-D images acquired in various locations are used in real-time to construct an image database with range information that can be used repetitively in future positioning applications. Details of image database construction, relative displacement calculation, SLAM state estimation process, and image comparison criteria for database image retrieval update are presented. Two different application scenarios are implemented to demonstrate the proposed method can perform SLAM in an unknown environment with simultaneous database build-up, and the RGB-D image database can be used in future applications even when conventional camera or point-cloud sensors are used as navigation sensor.