Monocular Depth Estimation with Self-Supervised Learning for Vineyard Unmanned Agricultural Vehicle

Area of Science:

• Computer Vision
• Robotics
• Agricultural Technology

Background:

• Accurate environmental perception is crucial for the safe and efficient operation of unmanned agricultural vehicles (UAVs).
• Traditional depth estimation methods often require expensive sensors or extensive labeled data, posing economic and practical challenges for UAV deployment.

Purpose of the Study:

• To develop an economical and efficient lightweight depth estimation model for UAVs using monocular cameras.
• To enable self-supervised learning for depth estimation, eliminating the need for depth information labels.

Main Methods:

• Proposed MonoDA, a convolutional neural network model comprising depth and pose estimation subnetworks.
• Utilized a modified U-Net for depth estimation and EfficientNet-B0 for pose estimation from sequential monocular frames.
• Implemented a self-supervised training strategy leveraging reprojection loss between adjacent frames and predicted poses.

Main Results:

• MonoDA demonstrated competitive accuracy on the KITTI raw dataset and a custom vineyard dataset.
• The model exhibited non-sensitivity to color variations, enhancing robustness in diverse agricultural environments.
• Achieved a processing speed of 18.92 FPS on an NVIDIA Jetson TX2, suitable for real-time UAV applications.

Conclusions:

• MonoDA offers an economical and effective solution for depth estimation in UAVs using monocular vision.
• The self-supervised approach significantly reduces data requirements and computational complexity.
• Presents a promising auxiliary depth detection paradigm for enhancing UAV navigation and environmental awareness.