John J. Leonard
Chryssostomos Chryssostomidis
Bradley A. Moran
Sea Grant College Prog., MIT, E38-300, 292 Main St., Cambridge, MA 02139
This talk will describe an approach to the recovery of geometric object shape descriptions from high-frequency sonar data for robotic applications. The goal is to recover explicit geometric surface descriptions for man-made objects, by fusing the geometric constraints of multiple sonar returns obtained by a moving autonomous underwater vehicle (AUV). Automated sensor data interpretation is made difficult by the pervasive problem of data association---uncertainty in the origins of measurements---in addition to the uncertainty in the values of measurements (noise) that is present in more traditional estimation and control problems. This problem is particularly severe when using sonar, for it is impossible to determine the geometry (e.g., sphere, cylinder, plane, corner, or spurious multiple reflection) that produced a single time-of-flight sonar return in isolation. One must first group measurements from different sensors and/or different sensing locations that have a common origin, to characterize the reflecting surface geometry. State estimation or surface fitting techniques can then be performed on correctly classified sets of measurements to complete the shape recovery process. Experimental results will be presented for two applications: (1) land robot map building and localization using the 50-kHz Polaroid ultrasonic ranging system, and (2) underwater scene reconstruction using a 1.25-MHz mechanically scanned profiling sonar.