Machine Vision for Urban Model Capture: Exploiting Scale, Achieving Automation
We describe a suite of scalable algorithms for acquiring calibrated imagery, and registering it to within a few centimeters and a fraction of a degree, over outdoor environments spanning hundreds of meters. Our sensor and associated algorithms form the foundation of a planned robotic 3D mapping capability in which one or more autonomous sensors will move about an architectural environment, capturing a 3D model of the environment. Researchers have developed a variety of techniques for inferring scene geometry and appearance from photographs. Historically, proposed techniques have been either algorithmic (but limited to small-scale, restricted settings), or interactive (so limited by the human operator's skill and capacity for work). To date, no single system has achieved both automation and scalable, end-to- end operation. We break the logjam by moving into an operating regime in which the scale of the problem becomes an advantage. Our algorithms assume that the urban scene to be captured exhibits collections of parallel lines, and identifiable window and building corners. We acquire thousands of images of the scene, and register and combine them using a suite of techniques incorporating geometric duality, geometric statistics, graph propagation, consensus techniques and numerical optimization. In contrast to the prevailing view that human intervention always improves quality, we demonstrate several tasks for which the automated system outperforms a human operator. We describe the status of the project and show some recent results.