DEVELOPMENT OF VISION SYSTEM AND END-EFFECTOR FOR AUTOMATIC BUD THINNING OF APPLE TREE: EARLY CROP LOAD MANAGEMENT
Restricted (Penn State Only)
- Author:
- Sahu, Rashmi
- Graduate Program:
- Agricultural and Biological Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 15, 2023
- Committee Members:
- Long He, Thesis Advisor/Co-Advisor
Paul Heinz Heinemann, Committee Member
Suat Irmak, Program Head/Chair
James Rawlinson Schupp, Committee Member
Shirin Ghatrehsamani, Committee Member - Keywords:
- : Computer Vision
YOLO
Robotic Bud Thinning
LED-Stereovision
Image Acquisition
Crop load management
End-effector
Apple flower bud thinning
Linear solenoid
Robotic bud thinning.
Computer Vision - Abstract:
- Crop load management is one of the most important production practices for apples to ensure good quality fruit for growers. Manual operations, such as pruning and thinning, are typically labor-intensive, expensive, and time-consuming. Therefore, robotic systems can be a potential solution to address the concerns of labor scarcity and high associated expenditures. Developing a robotic bud thinning system for early-stage crop load management in unstructured environments is crucial and demanding. This thesis focused on two core technologies for the development of a robotic bud thinning system for precision crop load management, e.g., a machine vision system for bud detection and an end-effector for bud removal. Real-time flower bud detection is the first step in developing a robotic bud thinning system. Detecting buds in natural orchard environments is a considerable challenge due to the complex environmental conditions, such as the tiny size of buds, different growth periods, and similarity between bud and cut branch parts. A rapid bud recognition method by using the deep learning, You Only Look Once (YOLO) model was used in this study. Three bud growth stages were used to identify flower buds for thinning, including silver tip, green tip, and tight cluster. The findings of this study demonstrated that the YOLOv4 outperformed the YOLOv5 and YOLOv7 models. YOLOv4 performed better in both one class (Bud) and three classes (silver tip, green tip, and tight cluster) groups. Meanwhile, the detection models' robustness was investigated using two datasets, e.g., images taken by stereo vision and mobile cameras. Additionally, the findings indicated that YOLOv4 outperformed YOLOv5 and YOLOv7 in terms of robustness across diverse datasets. The study provided an effective method of bud detection for robotic bud thinning. After developing an algorithm for bud identification, the next task focused on end-effector development for bud removal. Two end-effectors were developed, one using a scissor mechanism and the other with a rotating brush mechanism. Then the scissor and brush-type end-effectors were fabricated and evaluated in field tests for bud thinning. The field tests were conducted on both Gala and Fuji apple trees at three different bud stages (silver tip, green tip, and tight cluster), and 20 randomly selected branches for each test (one cultivar and one bud stage). The scissor end-effector is a solenoid actuated and was able to cut in less than 0.5 seconds once the bud was placed into the mechanism. The brush end-effector was operated by a motor and was capable of removing more accurately for small and lateral buds. In these field tests, both end-effectors achieved overall more than 90% success rates. Overall, the discoveries of this thesis are the first steps toward developing a bud-thinning robot that could lead to early crop load management for producing high-quality fruit by reducing the labor and chemical requirements in orchards. This robotic bud thinning method for crop load management includes the possibility of increasing the profitability for tree fruit growers and lowering environmental contamination.