Contributions to direct methods of estimation and control from visual data
Geraldo Silveira Filho · 2008
The overwhelming majority of vision-based techniques for both estimation and control consider a feature-based scheme. This thesis investigates how to appropriately exploit pixel intensities directly, i.e. without having to resort to image features. The fact of using all image information, even where no features exist, can considerably increase their accuracy and flexibility. To this end, we propose generic photo-geometric transformation models and optimization methods for directly and efficiently registering images (including color ones) of rigid and deformable objects, all in a unified manner. In particular, the new photometric model ensures robustness to arbitrary illumination changes, are independent of the object's attributes and of the camera's characteristics, and naturally encompasses gray-level images. We then show that the framework can effectively be formulated using uncalibrated or calibrated pinhole cameras. The differences mainly regard to the needed parametrization. A robust visual tracking technique is constructed by directly registering a reference image with successive frames. Then, using the optimal parameters that relate the two images, a vision-based control strategy is proposed to drive all six degrees-of-freedom of a robot to the (desired) pose where the reference image was taken. This new technique does not require either precise parameters of the vision system or any metric structure of the observed rigid scene, leading to a flexible and reliable system. If a calibrated camera is used, then the proposed robust visual tracking technique directly provides the optimal camera pose and scene structure. Since they are simultaneously and causally recovered, the technique represents a new solution to the visual Simultaneous Localization and Mapping (SLAM) problem. Finally, we propose a new visual servoing method that uses the estimates from this visual SLAM approach. Hence, this controlled visual SLAM scheme allows for autonomous navigation of mobile robots over previously unexplored scenes. Comparisons results with existing techniques demonstrate significant improvements in the system performance. Various real-world experiments and simulations are reported to show that the proposed methods can indeed be highly accurate and robust despite unknown objects and unknown imaging conditions. The trade-offs to attain real-time efficiency are discussed in the text.