Spectral Imaging for Greenhouse Cucumber Fruit Detection Based on Binocular Stereovision

Abstract

For a greenhouse cucumber-harvesting robot, two major challenges in the automated fruit-picking are to identify the target and to measure its position in the three dimensions. In this paper, a machine vision algorithm for recognition and location of cucumber fruits was presented. Firstly, a stereovision imaging system was established to capture monochrome near-infrared images, which are beneficial to deal with the similar-color segmentation problem in complex environment. Secondly, an image processing algorithm for cucumber recognition was proposed with the following steps: region partition according to 1-dimensional gray histogram distribution, adaptive threshold processing on divided local image, noise elimination using morphological analysis, and feature extraction with texture detection. Thirdly, the 3-D spatial position of fruit was calculated on the basis of standard triangulation model, meanwhile an auxiliary camera on the end-effector was introduced to improve cutting position. The experimental results from 120 cucumber image pairs taken in greenhouse showed that the proposed algorithm could detect fruit with a recognition rate of 86%. Also, the distance error of the grasping position(less than 8.6mm) and the maximum deviation of the cutting position (1.64mm in x-axis, 1.41mm in y-axis, 3.9mm in z-axis) were satisfied for robotic operation.