In-field citrus detection and localisation based on RGB-D image analysis. (October 2019)

• Record Type: Journal Article
• Authors: Lin, Guichao; Tang, Yunchao; Zou, Xiangjun; Li, Jinhui; Xiong, Juntao
• Is Part Of: Biosystems engineering. Volume 186(2019)
• Page Start: 34
• Page End: 44
• Publication Date: 2019-10
• Subjects: RGB-D image -- Bayes classifier -- Density clustering -- Support vector machine -- Citrus-harvesting robot; Bioengineering -- Periodicals; Agricultural engineering -- Periodicals; Biological systems -- Periodicals; Génie rural -- Périodiques; Systèmes biologiques -- Périodiques; 631
• DOI: 10.1016/j.biosystemseng.2019.06.019
• Languages: English
• ISSNs: 1537-5110

Abstract

Abstract : In-field citrus detection and localisation are highly challenging tasks due to varying illumination conditions, partial occlusion of citrus, and the colour variation of citrus at different stages of maturity. A reliable algorithm based on red-green-blue-depth (RGB-D) images was developed to detect and locate citrus in real, outdoor orchard environments for robotic harvesting. A depth filter and a Bayes-classifier-based image segmentation method were first developed to exclude as many backgrounds as possible. A density clustering method was then used to group adjacent points in the filtered RGB-D images into clusters, where each cluster represents a possible citrus. A colour, gradient, and geometry feature-based support vector machine classifier was trained to remove false positives. To test the method, a dataset with 506 RGB-D images was acquired in a citrus orchard on sunny and cloudy days. Results showed that the proposed algorithm was robust with an F1 score of 0.9197; the positioning errors in the x, y and z directions were 7.0 ± 2.5 mm, −4.0 ± 3.0 mm and 13.0 ± 3.0 mm, respectively, and the sizing error was −1.0 ± 4.0 mm. These excellent performance values demonstrate that the proposed method could be used to guide a citrus-harvesting robot. Highlights: An algorithm for citrus detection, localisation and sizing was developed. The algorithm can detect citrus robustly with an F1 score of 0.9197. Positioning errors in x, y, z were 7.0 ± 2.5 mm, −4.0 ± 3.0 mm and 13.0 ± 3.0 mm. The sizing error was −1.0 ± 4.0 mm.