Design and testing of a machine vision system for a robotic harvester of melons
Abstract
A machine vision system for a robotic melon harvester has been developed. The harvesting strategy and diverse field parameters have been analyzed, the image acquisition strategies designed, and the operation of the machine vision and image processing architecture developed. The architecture consists of three major components: the global frame analysis, the knowledge-directed evaluation, and the goal-oriented approach. The global frame analysis captures images of a melon bed section in its full width and identifies the melons in the image area. The knowledge-directed evaluation utilized the image parameters and the detection coordinates to separate multiple and error detections from correct melon detections. The goal-oriented approach guides the robotic manipulator during its approach toward the melon. Images are acquired and processed and the robotic controller provided with information about the distance to the melon and about deviations from the correct approach trajectory. Software prototypes have been developed on a Silicon Graphics Personal Iris for testing each level. The global analysis produces a large number of potential melons. The knowledge-based module then utilizes a database and a set of rules to reduce the list of possible melons to a high level of accuracy. The goal-oriented module then utilizes a triangulation algorithm to determine the distance to the melon and to identify an off-centered robotic approach. The system has been tested with field images of the melon beds generated under natural illumination conditions. The global frame analysis showed a detection performance of 86.5% to 89% detection accuracy (percentage of detected melons) under environmental conditions ranging from direct sunlight to clouded and dark illuminations. The images were acquired without artificial lighting or unnatural background. The disadvantage of this component was its low detection correctness. A high number of error detections and multiple detections of one melon were generated and would have reduced the efficiency of harvesting if not eliminated. The knowledge-directed evaluation eliminated over 90% of the error and multiple detections without significantly reducing the detection accuracy. The results show the applicability of machine vision and image processing to the melon harvesting operation and the feasibility of the proposed approach. The utilization of domain knowledge to evaluate the detection results of the conventional image processing software showed the potential of knowledge concepts in this domain. The goal-oriented sensing is required to support the robot operation in a natural environment. The concept was tested and showed its potential.
Degree
Ph.D.
Advisors
Miles, Purdue University.
Subject Area
Agricultural engineering|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.