Sensory perception for an agricultural robot

Meny Benady, Purdue University

Abstract

This thesis develops the sensory systems for an intelligent agricultural robot to selectively harvest ripe fruit, and a mathematical model for multi-sensor fusion that is used to quantitatively define fruit ripeness. The robot uses a structured light range scanner to detect fruits based on their three-dimensional shape and determines fruit ripeness by electronically sensing natural aromatic and nonaromatic gases emitted by the ripening fruit. Robotic melon harvesting has been undertaken as a case study. The sensor fusion approach developed is a general one for integration of multiple sensory inputs that are prone to inconsistencies or errors. The model was applied to the problem of determining fruit ripeness, by integrating multiple characteristic properties of ripening fruit, using a visual classification of the fruit into four ripeness classes as a classification variable. External color, flesh firmness, stem detachment force and TSS (Total Soluble Solids, which includes soluble sugars) served as the physical indicators of melon fruit ripening. The continuous ripeness values computed serve as a criteria for the evaluation of the robot's ripeness sensor. Fruit ripeness is determined by the robot using a semiconductor gas sensor that nondestructively measure changes in concentrations of aromatic and nonaromatic volatiles emitted by the ripening fruit. The sensor was generally more sensitive to fruit ripeness than other destructive and nondestructive methods, as evaluated in both laboratory tests, and in the field under varying ambient conditions. The response time of the sensor is less than 1 second. Fruit location is determined by using a laser range scanner to create cross-sections of field scenes. A four-valued labeling scheme is used to interpret the scene, with each range data point being labeled as UNKNOWN, MELON, PLANT or GROUND. The circular Hough transform in conjunction with domain specific knowledge based rules, are used for labeling the points and for consequent interpretation of the scene. An enhancement to the Hough transform was developed which regards not only the number of range points supporting detection of a curve in the scene, but also the distribution of the range points on the detected object's surface. This enables location of fruit despite significant occlusion by leaves. (Abstract shortened by UMI.)

Degree

Ph.D.

Advisors

Miles, Purdue University.

Subject Area

Agricultural engineering|Electrical engineering

COinS