Autonomous Visual Perception for Unmanned Surface Vehicle Navigation in an Unknown Environment.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous navigation of unmanned surface vehicles (USVs) requires reliable detection of water surfaces.
• Non-water regions pose significant risks as obstacles to USVs.
• Existing methods often rely on manual data labeling, which is time-consuming and labor-intensive.

Purpose of the Study:

• To propose a novel visual detection method for water regions essential for USV navigation.
• To develop an automated system that eliminates the need for manual data annotation.
• To enhance the robustness and adaptability of water detection algorithms for USVs.

Main Methods:

• An adaptive multistage segmentation algorithm clusters image pixels, assigning label tags and confidence values.
• A convolutional neural network (CNN) is trained online using the generated label and confidence maps.
• The online-trained CNN performs precise and robust segmentation of water regions.

Main Results:

• The proposed method successfully segments water regions in diverse lake environments.
• The system demonstrates robust performance in unknown navigation settings.
• The approach achieves high precision and reliability in water surface detection.

Conclusions:

• The developed method offers an efficient and automated solution for water region recognition in USV navigation.
• Online CNN training allows the system to adapt to varying environmental conditions.
• This technique significantly reduces the burden of manual data labeling for training deep learning models.