A Lightweight Forward-Looking Sonar Sensing Framework for Embedded Target Detection in Resource-Constrained...

Area of Science:

• Robotics and Autonomous Systems
• Marine Technology
• Computer Vision

Background:

• Forward-looking sonar (FLS) is crucial for underwater navigation in low-visibility conditions.
• Target detection in FLS data is difficult due to weak echoes, clutter, and computational constraints.
• Existing methods struggle with the resource limitations of embedded marine robotic systems.

Purpose of the Study:

• To develop a lightweight and efficient FLS sensing framework for embedded target detection.
• To improve target detection accuracy and computational efficiency in resource-constrained underwater systems.
• To enable reliable operation in challenging marine environments.

Main Methods:

• A compact FPN-Mix backbone with a Conv-Mix module for efficient feature aggregation.
• Difficulty-aware supervision using a target-aware dynamic weighting loss for challenging samples.
• Multi-level knowledge distillation from a teacher model to a compact student detector.
• Model optimization including ONNX export and TensorRT FP16 acceleration.

Main Results:

• The proposed framework achieves a balance between detection accuracy and computational efficiency.
• The model demonstrates competitive performance on public and field datasets.
• The optimized model achieves 72.23 FPS on an NVIDIA Jetson Orin NX platform.
• The final model has only 2.83 M parameters and requires 6.68 GFLOPs.

Conclusions:

• The lightweight FLS sensing framework is suitable for embedded target detection in underwater systems.
• The method effectively addresses challenges of weak echoes, clutter, and computational limits.
• The system's efficiency and accuracy support practical deployment in real-world marine applications.