ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

RGB-D Perception for Selective Harvesting of Field-Grown Asparagus with Dual-Path Detection and Point-Cloud-Free Measurement

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600512.(doi:10.13031/aim.202600512)
Authors:   Dongyu Liu, Yuzhen Lu
Keywords:   Artificial intelligence; Depth image; Diameter estimation; Green asparagus; RGB-D perception; Selective harvesting.

Abstract. Green asparagus is a labor-intensive crop, with harvesting alone accounting for up to 60 % of total production costs in the United States. Despite the need for harvest automation, commercially viable selective-harvesting systems for open-field conditions remain unavailable. While previous studies have demonstrated the feasibility of selective harvesting under laboratory or controlled environments, reliable spear perception and measurement under production field conditions remain major technical challenges. This study presents an innovative RGB-D perception pipeline for selective harvesting of field-grown green asparagus, integrating spear detection, length and diameter estimation, and base point localization. First, a dual-path detection model augments a pretrained YOLOv12s backbone with a lightweight Depth Encoder that injects depth features at multiple feature pyramid scales. Trained on a two-season dataset, the model achieved mAP50 of 80.68 %, a 4.9% improvement over the RGB-only baseline under challenging field variability. Second, a point-cloud-free measurement pipeline estimates spear base-point location, length, and diameter directly from aligned RGB-D image pairs. On true positive detections, the pipeline achieved a length estimation accuracy of 90.8% with the corresponding mean absolute error (MAE) of 2.64 cm, diameter estimation accuracy of 86.7% (MAE=2.82 mm), and a mean spear base point localization error of 2.20 cm (median=1.39 cm). Under the joint length-diameter-localization criteria, the system achieved an overall classification accuracy of 86.5%, with a recall rate of 92.1% for harvestable spears on true positive detections, reflecting deployment-level performance. The end-to-end unified pipeline processes each spear in approximately 18 ms, and in particular, the measurement component achieved a 3-fold reduction in per-spear inference time compared with a state-of-the-art point-cloud-based approach while substantially improving coverage. Overall, the proposed system provides a robust and efficient perception foundation for selective harvesting of green asparagus in production field environments.

(Download PDF)    (Export to EndNotes)