Real-time pose estimation for an underwater object combined with deep learning and prior information

Area of Science:

• Robotics and Computer Vision
• Underwater Imaging and Sensing

Background:

• Underwater autonomous operations using monocular vision suffer from low accuracy and intelligence due to imprecise pose estimation.
• Image degradation and optical refraction present significant challenges for underwater visual perception.

Purpose of the Study:

• To develop a real-time, high-accuracy pose estimation method for underwater cylinders and cuboids using monocular vision.
• To address image degradation and optical refraction challenges in underwater environments.

Main Methods:

• Image dehazing using a scale-optimized dark channel prior algorithm.
• Object detection and pixel information extraction using a lightweight, improved You Only Look Once v5 (YOLOv5).
• Pose estimation leveraging an underwater optical imaging model and an improved perspective-n-point algorithm.

Main Results:

• The proposed method achieves accurate real-time pose estimation for underwater objects.
• The algorithm demonstrates excellent performance when deployed on edge computing devices like NVIDIA Jetson TX2.
• High-precision monocular pose estimation is achieved without the need for large-scale pose datasets.

Conclusions:

• The developed method significantly improves the accuracy and intelligence of underwater autonomous operations.
• This technique provides reliable pose information crucial for navigation and task execution in underwater robotics.
• The approach offers a practical solution for enhancing monocular vision-based underwater systems.