Vision Systems for the Location of Citrus-Fruit in a Tree Canopy

Abstract

One of the main factors for the automation of citrus harvesting to be made economically viable is to achieve an acceptably efficient vision system. For this purpose, algorithms capable of identifying and locating the greatest amount of fruits must be designed, which minimize the rate of misclassified objects.

In this work, a study of changes of the reflection spectrum of the peel of fruits and leaves, has been conducted during the harvesting period. With these data, two algorithms for the detection of fruit have been developed; one based on thresholding of an image taken with a red filter, and the other using information from two images filtered in the red and green bands respectively. An additional solution has been investigated, based on the study of the object light reflectance, through the curvature of the image surface. It was demonstrated that the use of flashes during image acquisition improved results. About 80% of visible fruit in the tree was detected with the algorithms implemented, with misclassification rates not exceeding 10%. The method based on the red/green relationship, proved capable of identifying the largest percentage of fruit, whereas the curvature method was the most precise.