Monocular and range camera cross-calibration for RGB-D sensor architectures

摘要

RGB-D sensor frameworks such as the PrimeSense/Kinect have brought a massive change in the range of applications for the usage of depth data in not just core robotic and computer vision systems, but also in security, entertainment and medical faculties among others. Such projected texture range measurement systems have also effectively substituted traditional range sensor systems such as Laser and Lidar, which are not just bulky and expensive, but offer poor resolution/unit cost and low speed of usage. On the other hand, generic RGB-D sensor frameworks (as opposed to integrated RGB-D cameras) that provide flexibility in terms of usage of variegated monocular color and range image sensors form the future of computer vision applications. Unlike fixed RGB-D frameworks, these generic frameworks require explicit cross-calibration between the range and the monocular color image sensors. Traditional 2D checkerboard or similar alternate calibration patterns do not provide the necessary sensory response across the varied sensing modalities for accurate cross-calibration. To address this concern, we present a novel framework for extrinsic cross-calibration of variegated monocular and range sensors by extension of the traditional checkerboard pattern used for monocular or stereo calibration into a 3D checkerboard framework. A suite of computer vision techniques are also presented in order to obtain the necessary calibration parameters using the presented calibration pattern. Results presented show successful detection of correspondence points and estimation of extrinsic parameters for cross-calibration. It can also be seen that the error in the system increases with depth as the estimates from the Kinect sensor become unreliable.