Real-Time Pipe and Valve Characterisation and Mapping for Autonomous Underwater Intervention Tasks

Area of Science:

• Robotics and Automation
• Computer Vision
• Marine Engineering

Background:

• Underwater operations for infrastructure inspection are increasingly complex and risky.
• Autonomous Underwater Vehicles (AUVs) offer a solution for automating these tasks, reducing risk and time.
• Vision-based sensing, particularly RGB data, is crucial for detailed underwater inspection and manipulation.

Purpose of the Study:

• To develop and validate a deep neural network for pixel-wise 3D segmentation of underwater pipes and valves.
• To create algorithms for extracting critical information from segmented underwater infrastructure.
• To demonstrate the real-time applicability of the system on an AUV for practical subsea operations.

Main Methods:

• Utilized a deep neural network for 3D point cloud segmentation of pipes and valves from stereo camera data.
• Developed novel algorithms for extracting pipe vectors, gripping points, structural elements, and valve information.
• Implemented and tested the system in real-time on an Autonomous Underwater Vehicle.

Main Results:

• Achieved high segmentation performance with a mean F1-score of 88.0% (pixel-wise) and 95.3% (instance-wise).
• Demonstrated excellent accuracy in extracting information from pipe instances and good performance for valves.
• Validated real-time execution on an AUV with a frame rate of 0.72 fps, suitable for manipulation and inspection.

Conclusions:

• The developed deep learning approach enables accurate 3D segmentation and information extraction for underwater infrastructure.
• The system's real-time performance on AUVs supports practical applications in subsea inspection and manipulation.
• The provided dataset, model, and algorithms contribute to advancing underwater robotics research.