Stereo Vision for Plant Detection in Dense Scenes.

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Automated weed control relies on accurate crop and weed visual discrimination.
• Current deep learning models struggle with weed detection in dense, occluded agricultural scenes.
• The integration of color (RGB) and depth (D) data in deep learning for agriculture is underutilized.

Purpose of the Study:

• To evaluate the effectiveness of combining RGB and D data for plant detection in agriculture.
• To introduce and assess novel deep learning models utilizing RGB-D data.
• To investigate the impact of data augmentation and transfer learning on RGB-D plant detection models.

Main Methods:

• Collected a novel RGB-D dataset of sugar beet crops with varying weed densities.
• Developed D-model (depth data) and CD-model (color and depth data) for plant detection.
• Transformed depth data into color-encoded 2D images for compatibility with existing architectures.
• Evaluated models using data augmentation and transfer learning techniques.

Main Results:

• Geometric data augmentation and transfer learning proved effective for both RGB and RGB-D models.
• The CD-model demonstrated improved performance by integrating color-encoded depth data.
• Despite improvements, depth data did not enhance volunteer potato detection in dense sugar beet fields.

Conclusions:

• Combining color-encoded depth data with data augmentation and transfer learning can enhance plant detection models.
• The utility of depth data for weed detection is context-dependent, especially in high-density vegetation.
• Further research is needed to optimize RGB-D fusion for challenging agricultural scenarios.