Dynamic stereo vision

摘要

Sensing 3-D shape and motion is an important problem in autonomous navigation and manipulation. Stereo vision is an attractive approach to this problem in several domains. In this thesis, I address fundamental components of this problem by using stereo vision to estimate the 3-D structure of of objects visible to a robot, as well as to estimate the motion of the robot as its travels through an unknown environment. I begin by using cameras on-board a robot vehicle to estimate the motion of the vehicle by tracking 3-D feature-points or landmarks. I formulate this task as a statistical estimation problem, develop sequential methods for estimating the vehicle motion and updating the landmark model, and implement a system that successfully tracks landmarks through stereo image sequences. In laboratory experiments, this system has achieved an accuracy of 2$\\$% of distance over 5.5 meters and 55 stereo image pairs. These results establish the importance of statistical modelling in this problem and demonstrate the feasibility of visual motion estimation in unknown environments. This work embodies a successful paradigm for feature-based and motion estimation, but the feature-based approach results in a very limited 3-D model of the environment. To extend this aspect of the system, I address the problem of estimating depth from stereo images. Depth maps specify scene for each pixel in the image. I propose a system architecture in which exploratory camera motion is used to acquire a narrow-baseline image pair by moving one camera of the stereo system. Depth estimates obtained from this image pair are used to matching of a wide-baseline image pair acquired with both cameras of the system. I formulate the bootstrap operation statistically by modelling maps as random fields and developing Bayesian matching algorithms in which information from the narrow-baseline image pair determines the prior density for matching the wide baseline image pair. This leads to efficient, area-based matching algorithms that are applied independently for each pixel or each scanline of the image. Experimental results with images of complex, outdoor scene models demonstrate the power of the approach.