Dynamic detection of three-dimensional crop phenotypes based on a consumer-grade RGB-D camera

Area of Science:

• Agricultural Science
• Computer Vision
• Robotics

Background:

• Nondestructive field-based crop phenotyping is crucial for crop breeding.
• Ground-based mobile platforms with sensors offer efficient and accurate data collection.

Purpose of the Study:

• To propose a dynamic 3D data acquisition method for various crops using a consumer-grade RGB-D camera on a mobile platform.
• To enable efficient and accurate measurement of 3D phenotypic traits for individual plants.

Main Methods:

• Utilized a scale-invariant feature transform (SIFT) operator for coarse point cloud alignment.
• Employed the colored iterative closest point (ICP) algorithm for fine matching and 3D point cloud generation.
• Applied clustering to segment individual plants and measure 3D traits like plant height, leaf area, and projected area.

Main Results:

• Demonstrated strong correlations between measured and manual results for plant height (R²=0.9–0.96), leaf area (R²=0.8–0.86), and projected area (R²=0.96–0.99).
• Validated the method's applicability across different crops (corn, tobacco, cotton, Bletilla striata) at the seedling stage.
• Confirmed successful dynamic detection at speeds up to 0.6 m/s, with acceptable results during day and night.

Conclusions:

• The proposed method enhances the efficiency of 3D point cloud data extraction for individual crops with acceptable accuracy.
• This dynamic 3D data acquisition technique presents a feasible solution for outdoor crop seedling phenotyping.