Perception Algorithms for a Harvesting Robot

Abstract

Lawn mowing, snow removal, and alfalfa harvesting are all surface-covering tasks; that is, members of a class of tasks which require a vehicle to be guided over a surface while processing the surface in some way. The planning community has recognized for some time that surface covering tasks can be usefully addressed as a general category. We believe that this viewpoint is useful for perception as well; in particular, we claim that surface covering tasks share both a common set of perception needs and some common domain knowledge that can be exploited to meet those needs. To substantiate this claim, we have developed methods for addressingthree different surface covering perception tasks: boundary sensing, end-of-row detection, and obstacle detection. The boundary sensing problem is explored in some depth, using three different original algorithms: a simple model-fitting method, a curvature postulation method derived from the road-following system RALPH, and a method based on Bayesian statistics. Using exam-ples from the Demeter automated harvester project, we demonstrate the advantages of adaptive algorithms over static ones and discuss several methods for implementing an adaptive capability. As a sideline, a partial solution is offered for detecting and compensating for shadows which would otherwise disrupt the algorithms. This work is demonstrated on images from several surface-covering domains: alfalfa harvesting by color segmentation, snow removal, by brightness discrimination, and trash com-paction by texture segmentation. The most significant contribution, however, is the development of a commercially viable guidance system for the Demeter harvester; as of this writing, over 80 acres of alfalfa have been harvested autonomously from fields in Kan-sas, California, and Pennsylvania. New Holland has plans to offer the vision system as part of a commercially available product for sale in the year 2000.