Calibration and Multipath Mitigation for Increased Accuracy of Time-of-Flight Camera Measurements in Robotic Applications

This dissertation entitled "Calibration and Multipath Mitigation for Increased Ac- curacy of Time-of-Flight Camera Measurements in Robotic Applications" discusses the systematic measurement errors of a Time-of-Flight (ToF) camera and introduces two approaches to compensate for these errors. ToF cameras are active, imaging sensors, which provide depth images and monochromatic images with a frame rate up to 50 Hz. These devices allow not only for two-dimensional computer vision algorithms but also for three-dimensional perception and, thus, they arouse great interest in robotics. This work addresses the characteristic measurement errors and designs an appropriate error model. It focuses on two classes of errors, which are considered independently of each other. First, the typical, distal intrinsic ToF measurement errors are described. These are caused by the imperfect implementation of the measurement process. Second, the multipath reflections generated outside of the camera are analysed. To begin with, the dissertation designs a ToF camera error model and a corresponding calibration procedure that identifies the so-called distal intrinsic parameters. The ToF camera naturally is an imaging sensor and attached to the tool- centre-point of the robot. Besides the distal parameters, therefore, also the lateral intrinsic and the extrinsic parameters have to be identified in several steps. For this reason, the dissertation explores a holistic calibration procedure that calculates all parameters in one step - solely based on the depth images and managed without image feature extraction. Multipath depends on the vicinity of the sensor and cannot be characterised in advance. The infrared light unhamperedly propagates in the scene, hence, the cam- era pixels receive light that is reflected multiple times. This immanent phenomenon causes errors in the range of several centimetres. The dissertation explains the origination of multipath by means of an uncomplex model. For this purpose, the signal paths of the individual camera pixels are traced back similar to radiosity and ray-tracing approaches. This procedure allows for a calculation of the additional, interfering signal component and, thus, for a correction of the measurement. The improvement factor of the measurement accuracy thanks the two approaches is investigated experimentally. The usage of three distinct ToF devices accentuates the general validity of the approaches. The calibration outcomes demonstrate that the ToF error model is suited to compensate for the systematic errors. The one- step calibration procedure, however, is inferior to the common photogrammetric methods in terms of robustness and accuracy. In total, the accuracy of the robotic ToF camera system is better than 10 mm and 0.5 deg. The successful compensation for the multipath-related error by as much as 80 % is proved by means of ego- motion estimation experiments. Finally, a Bin-Picking application is presented, which provides evidence for the practicality of the ToF camera calibration