Image/Map Correspondence Using Curve Matching†

We address the general practical problem of determining correspondences between maps and terrain images, and focus on a static low altitude airborne scenario. For this case we consider the approach of partially matching detected and expected curves in the image plane. Expected curves are generated from a map, using an estimate of the sensor pose in three dimensions, and matched with simulated detected curves in an image. We also outline a method for sensor pose refinement using point correspondences derived from curve matches as input to a relative orientation algorithm. Image/Map Correspondence We address the problem of determining correspondences between maps and terrain images in low altitude airborne scenarios. In particular, we consider an aircraft with a wide-angle passive imaging sensor, such as an infrared or TV camera, flying a few hundred feet above tree tops. Other on-board systems are assumed to provide an estimate of the three-dimensional orientation of the imaging sensor. We also assume availability of an estimate of three-dimensional position, for which there are several sources of information in varying accuracies (e.g. altimeters, inertial navigation, and global positioning systems). The important point is that regardless of how accurate the sensor pose estimate is in principle, it will be inaccurate in detail in practice. Hence we seek effective means of determining correspondences between images and maps, to provide enhanced relative direction and absolute position information. Finally, we assume availability of a variety of maps for use by the system. The maps must be complete enough to permit prediction of visually prominent aspects of the environment. For this a terrain elevation map is critical but not sufficient. Other maps outlining the locations of areas such as forests, fields, water, roads, and buildings are also needed. Solutions to the image/map correspondence problem are important for visual support of navigation, and critical in guiding the search for objects in areas anticipated from map information. Image/map correspondence is also important in low altitude airborne surveillance and reconnaissance problems, in which the objective is to refine existing map information on the basis of incoming imagery. Examples of other work directly addressing various formulations of passive image/map correspondence problems are: (Bolles et al. 1979), (Medioni & Nevatia 1984), (Mokhtarian & Mackworth 1986), (McKeown 1987), (Lawton et al. 1987), and (Andress & Kak 1988). Our overall approach is conceptually most similar to the approach of Mokhtarian & Mackworth (1986). In contrast to these prior efforts, we emphasize the low altitude formulation of the problem and the roles of partial curve matching and expected curve generation. The sensor position considered in this paper, at low altitudes above tree tops, poses significantly different problems and considerations than those encountered in land vehicle vision systems. For example, from viewpoints on the ground, forests are significant obstacles to vision and can easily obscure vast amounts of information such as the contour of the horizon. In this case a more map-like scale of visual information is available from the air for comparison with maps because only areas in the immediate vicinity of trees are occluded. The assumed low altitude sensor position also poses significantly different problems than encountered in high altitude aerial and satellite imagery. In the low altitude case, useful viewing directions tend more toward the horizon than straight down, forcing a threedimensional formulation of the solution, as opposed to the largely two-dimensional approaches that are viable at high altitudes. Of course the substantial body of work in modelbased object recognition, e.g. as surveyed by Besl & Jain (1985), is also relevant in principle. In fact, the