CAM-UNet: A Novel Water Environment Perception Method Integrating CoAtNet Structure

Area of Science:

• Computer Vision
• Robotics
• Marine Technology

Background:

• Accurate segmentation of aquatic environments is vital for unmanned surface vessel (USV) navigation.
• Existing models struggle with complex water textures, blurred boundaries, and capturing both global context and fine spatial details, leading to fragmented results.

Purpose of the Study:

• To develop a novel segmentation model that accurately identifies navigable waters and obstacles in challenging aquatic environments.
• To improve the performance of USV navigation systems by enhancing segmentation accuracy and robustness.

Main Methods:

• A novel segmentation framework based on the CoAtNet architecture was developed.
• The framework features an enhanced convolutional attention encoder with Fused-MBConv and CBAM modules for refined boundary and feature awareness.
• A Bi-level Former (BiFormer) and Multi-scale Attention Aggregation (MSAA) module were integrated for collaborative global-local feature modeling and multi-scale contextual information capture.
• A U-Net-based decoder was employed for gradual spatial resolution restoration.

Main Results:

• The proposed model achieved 95.15% mean Intersection over Union (mIoU) on a self-collected dataset.
• The model attained 98.48% mIoU on the public MaSTr1325 dataset.
• Performance surpassed established models including DeepLabV3+, SeaFormer, and WaSRNet.

Conclusions:

• The novel segmentation model demonstrates superior performance in interpreting complex aquatic environments for autonomous navigation.
• The integrated approach effectively addresses limitations of current models in handling intricate water textures and boundary ambiguities.
• The model provides a robust solution for accurate navigable water and obstacle segmentation, crucial for safe and efficient USV operations.