Mushroom Detection and Three Dimensional Pose Estimation from Multi-View Point Clouds

Area of Science:

• Robotics
• Computer Vision
• Agricultural Science

Background:

• Agricultural robotics is crucial for efficient farm tasks.
• Mushroom harvesting requires precise robotic manipulation.
• Accurate 3D pose estimation is vital for robotic perception.

Purpose of the Study:

• To develop a vision module for 3D mushroom pose estimation.
• To address the challenge of limited 3D annotation data in agricultural settings.
• To enable precise robotic harvesting in industrial mushroom farms.

Main Methods:

• Utilized multi-view point clouds from RealSense active-stereo cameras.
• Developed a novel pipeline for mushroom instance segmentation and template matching.
• Employed a 3D mushroom model as the sole required data input.

Main Results:

• Successfully estimated the 3D pose of mushrooms from point cloud data.
• Demonstrated effectiveness despite the lack of large-scale 3D annotation.
• Validated the approach on both synthetic and real-world mushroom datasets.

Conclusions:

• The developed vision module enables accurate 3D mushroom pose estimation.
• The novel pipeline effectively overcomes annotation data limitations.
• This research advances the capabilities of agricultural robots for harvesting tasks.