In this paper a system capable of obtaining the 3D pose of a mobile robot using an external calibrated camera is proposed. The system robustly tracks point fiducials in the image plane generated by the robot's rigid shape in motion. Each fiducial is identified with a point belonging to a sparse 3D geometrical model of robot's structure. Such model allows direct pose estimation from image measurements and it can be easily enriched at each iteration with new points as the robot motion evolves. The entire process is solved online by using recursive Bayesian inference of the present pose given the measurements. The approach allows to model properly uncertainty in measurements and estimations, at the same time it serves as a regularization step in pose estimation. Initialization is solved by using odometry information available in the robot, jointly with image measurements to generate a geometrical model. A bundle-adjustment technique is used to properly model odometry drift error. The proposed approach is verified using simulated and real data.
展开▼
