Few-Shot Object Detection Using Multimodal Sensor Systems of Unmanned Surface Vehicles

Area of Science:

• Robotics
• Computer Vision
• Marine Technology

Background:

• Object detection is crucial for autonomous unmanned surface vehicles (USVs).
• Complex marine conditions and limited labeled data hinder traditional object detection performance.
• Shipborne sensors face interference, impacting detection accuracy.

Purpose of the Study:

• To propose a few-shot surface object detection method for USVs using multimodal sensor systems.
• To enhance the detection capabilities of USVs in challenging marine environments.
• To address the limitations of traditional object detection with insufficient labeled data.

Main Methods:

• Developed a few-shot object detection approach leveraging multimodal sensor fusion.
• Implemented metric learning for enhanced few-shot object detection on USVs.
• Utilized three-dimensional object detection with multimodal sensors to improve moving object detection.

Main Results:

• The proposed method significantly improved classification results for few-shot tasks.
• Achieved performance gains of 2-10% in average precision (AP) across datasets.
• Demonstrated improvements of 24-28% in average orientation similarity (AOS).

Conclusions:

• The multimodal, few-shot object detection method enhances USV capabilities in data-scarce scenarios.
• This approach offers a viable solution for applications with limited labeled datasets.
• The study provides a foundation for more robust autonomous marine systems.