Eye-to-eye calibration: extrinsic calibration of multi-camera systems using hand-eye calibration methods

The problem addressed in this thesis is the extrinsic calibration of embedded multi-camera systems without overlapping views, i.e., to determine the positions and orientations of rigidly coupled cameras with respect to a common coordinate frame from captured images. Such camera systems are of increasing interest for computer vision applications due to their large combined field of view, providing practical use for visual navigation and 3d scene reconstruction. However, in order to propagate observations from one camera to another, the parameters of the coordinate transformation between both cameras have to be determined accurately. Classical methods for extrinsic camera calibration relying on spatial correspondences between images cannot be applied here. The central topic of this work is an analysis of methods based on hand-eye calibration that exploit constraints of rigidly coupled motions to solve this problem from visual camera ego-motion estimation only, without need for additional sensors for pose tracking such as inertial measurement units or vehicle odometry. The resulting extrinsic calibration methods are referred to as "eye-to-eye calibration". We provide solutions based on pose measurements (geometric eye-to-eye calibration), decoupling the actual pose estimation from the extrinsic calibration, and solutions based on images measurements (visual eye-to-eye calibration), integrating both steps within a general Structure from Motion framework. Specific solutions are also proposed for critical motion configurations such as planar motion which often occurs in vehicle-based applications.