An "intelligent" system must be able to sense, plan, and control. It has become increasingly evident that only when these capabilities work together a system can successfully accomplish non-trivial tasks. In this thesis, we discuss issues that are related to such systems in the context of a vision-based quadruped walking robot system that we have designed and developed.; The progress made so far in the design of legged robots has dealt mostly with the issues of leg coordination, gait control, stability, incorporation of various types of sensors, etc. This progress has resulted in the demonstration of rudimentary robotic walking capabilities in various labs around the world. The more stable of these robots have multiple legs, four or more, and some can even climb stairs. But what is missing in most of these robots is some perception-based high-level control that would permit a robot to operate intelligently. Equipping a robot with perception-based control is not merely a matter of adding to the robot yet another module; the high-level control must be tightly integrated with the low-level control needed for locomotion and stability. For our high-level control, a model-based method to recognize a staircase using a single 2D image of a 3D scene is studied. The staircase recognition is achieved by obtaining the pose of a camera coordinate frame which aligns model edges with image edges. The method contains the matching, the pose estimation, and the refinement procedures. We propose a new matching scheme to reduce the complexity of correspondence search between model and image features. This is accomplished by grouping edges with certain geometric characteristics together. The refinement process uses all matched features to tightly fit the model edges with camera image edges. The resulting recognition is used to guide the robot to climb stairs.
展开▼
