Guava Detection and Pose Estimation Using a Low-Cost RGB-D Sensor in the Field

Area of Science:

• Robotics
• Computer Vision
• Agricultural Engineering

Background:

• Automatic harvesting requires precise fruit detection and pose estimation for collision-free robotic operation.
• Guava harvesting necessitates understanding fruit orientation relative to branches for safe detachment.

Purpose of the Study:

• To develop and evaluate a method for fruit detection and 3D pose estimation using a low-cost RGB-D sensor for automatic guava harvesting.
• To enable collision-free robotic picking by accurately determining fruit position and orientation with respect to branches.

Main Methods:

• Utilized a fully convolutional network for fruit and branch segmentation from RGB images.
• Applied Euclidean clustering on RGB-D data for individual fruit point cloud grouping.
• Developed a 3D line-segment detection method for branch reconstruction.
• Estimated fruit 3D pose using centroid and nearest branch information.

Main Results:

• Achieved high precision (0.983) and recall (0.948) in guava fruit detection.
• Reported a 3D pose error of 23.43° ± 14.18°.
• Demonstrated an execution time of 0.565 seconds per fruit.

Conclusions:

• The proposed method effectively detects guava fruits and estimates their 3D pose in outdoor conditions.
• The system is suitable for integration into a guava-harvesting robot, facilitating automated and collision-free operations.