A non-parametric Calibration Algorithm for Depth Sensors Exploiting RGB Cameras

Range sensors are common devices on modern robotic platforms. They endow the robot with information about distance and shape of the objects in the sensors field of view. In particular, the advent in the last few years of consumer RGB- D sensors such as the Microsoft Kinect, has greatly fostered the development of depth-based algorithms for robotics. In fact, such sensors can provide a large quantity of data at a relatively low price. In this thesis three different calibration problems for depth sensors are tackled. The first original contribution to the state of the art is an algorithm to recover the axis of rotation of a 2D laser range finder (LRF) mounted on a rotating support. The key difference with other approaches is the use of kinematics point-plane constraints to estimate the pose of the LRF with respect to a static camera, and screw decomposition to recover the axis of rotation. The correct reconstruction of a small indoor environment after calibration validates the proposed algorithm. The second and most important original contribution of the thesis is a fully automatic two-steps calibration algorithm for structured-light depth sensors (e.g. Kinect). The key novelty of this work is the separation of the depth error into two components, corrected with functions estimated on a pixel-basis. This separation, validated by experimental observations, allows to dramatically reduce the number of parameters in the final non-linear minimization and, consequently, the time for the solution to converge to the global minimum. The depth images of a test set corrected using the obtained calibration parameters are analyzed and compared to the ground truth. The comparison shows that they differ from the real ones just for an unpredictable noise. A qualitative analysis of the fusion between depth and RGB data further confirms the effectiveness of the approach. Moreover, a ROS package for both calibrating and correcting the Kinect data has been released as open source. The third contribution reported in the thesis is a new distributed calibration algorithm for networks composed by cameras and already-calibrated depth sensors. A ROS package implementing the proposed approach has been developed and is available for free as a part of a big open source project for people tracking: OpenPTrack. The developed package is able to calibrate networks composed by a dozen sensors in real-time (i.e., batch processing is not needed), exploiting plane- to-plane constraints and non-linear least squares optimization.