Visual motion processing for active observers

Computer vision will play a major role in future mobile robotics. This is due to the high resolution of sensor input and the variety of relevant tasks to solve such as motion estimation, distance computation and object recognition. Along with the development of algorithms it is necessary to consider their robustness as well as their real time capability in order to enable the operation in real world environments. This dissertation introduces three innovative solutions based on image sequences for one of the above mentioned basic capabilities: the monocular estimation of observer motion w.r.t. a scene. Extensive testing results with simulation tools and real world image sequences are presented. Robustness and real time issues are discussed on the basis of these results. The solutions are compared to outline their benefits and drawbacks. The first solution computers the translational direction of a camera on the basis of global information using a suitable cost function and minimizing by gradient descent. The solution is tested on real robotic hardware and proved to be very robust towards various types of statistical and systmatic noise. The second solution establishes a convergent approach based on a novel key-point detection scheme for the estimation of the translational egomotion state of the observer. Paraboloid cost functions guarantee convergence. The solution is simulated and tested with synthetic as well as real sequences indicating the robustness of the approach. The third solution presents a local method for the high resolution estimation of the relative general motion (i.e., translational as well as rotational) by linear transformations and a statistically advantageous local processing. The concept is based on achieving measurement independence for robust matrix inversion by suitable modulation and statistical averaging. A novel method for the efficient estimation of model-parameters from measured data is presented. The approach is compared with various traditional alternatives. The extensions w.r.t. the first two solutions concern the consideration of dynamic environments, i.e., other independently moving objects, and the joint estimation of translation as well as rotation. Testing indicates the expected tradeoff between local operation and noise sensitivity on one hand and global operation but robustness on the other hand. Simulations are performed with the newly developed 3D simulation tool Simon. The tool provides images, depth maps analytically correct spatiotemporal intensity derivatives for error analysis based on the camera motion state and a planar surface to be freely placed in 3D space (orig.)Available from TIB Hannover: RR 7240(96,1) / FIZ - Fachinformationszzentrum Karlsruhe / TIB - Technische InformationsbibliothekSIGLEDEGerman