WaSR-A Water Segmentation and Refinement Maritime Obstacle Detection Network.

Area of Science:

• Computer Vision
• Robotics
• Marine Engineering

Background:

• Semantic segmentation is crucial for autonomous vehicle obstacle detection.
• Existing methods struggle in aquatic environments due to water reflections and wakes, causing false positives.
• There is a need for specialized segmentation models for marine applications.

Purpose of the Study:

• To develop a novel deep learning architecture for robust water segmentation in marine environments.
• To improve the accuracy of obstacle detection for unmanned surface vehicles (USVs).
• To reduce false positive detections caused by visual ambiguities like water reflections and fog.

Main Methods:

• Proposed a novel deep encoder-decoder architecture named Water Segmentation and Refinement (WaSR) network.
• Utilized a ResNet101 encoder with atrous convolutions for feature extraction.
• Integrated inertial measurement unit (IMU) data into the decoder to enhance segmentation accuracy.
• Introduced a novel loss function for semantic separation to improve robustness.

Main Results:

• WaSR significantly reduced false positives (FPs) and increased true positives (TPs).
• Achieved approximately 4% higher F1 score compared to state-of-the-art methods on a USV dataset.
• Demonstrated strong generalization capabilities, outperforming state-of-the-art by over 24% in F1 score on a domain generalization experiment.

Conclusions:

• The WaSR network effectively addresses the challenges of semantic segmentation in marine environments.
• Integration of IMU data and a novel loss function enhances segmentation accuracy and robustness.
• WaSR represents a significant advancement for obstacle detection in autonomous marine systems.