Application of Machine Vision for Automated Weeding and Thinning in Broadcast-Grown Vegetables
作者
Jyh‐Cheng Yu,Hsueh-Yen Min,Chihpin Huang,Chih-En Lin
标识
DOI:10.1109/irce66030.2025.11202979
摘要
This paper presents a machine vision-based robotic system for automated weeding and thinning of short-cycle vegetables grown via broadcast seeding, a common practice in Taiwanese agriculture. A YOLOv8 object detection model was trained using 1,153 augmented images of 12-day-old Baby Bok Choy fields, achieving an average precision of 89.6% for crops and 77.7% for weeds. To improve localization accuracy, plant contour centroids were extracted from bounding boxes using image processing techniques. A thinning algorithm based on Poisson Disk Sampling (PDS) was employed to determine which seedlings should be retained or removed. A gantry-type robot prototype was developed, integrating a Delta robotic arm for weeding and thinning operations. RGB-D data from Intel RealSense D455 cameras was fused with ORB-SLAM3 to convert image coordinates to robot-space 3D coordinates. Experimental validation in a 4-meter greenhouse plot demonstrated the system's ability to accurately identify and execute crop management tasks while generating a crop distribution map. This map enables improved yield estimation and supports downstream precision farming applications.