eWaSR-An Embedded-Compute-Ready Maritime Obstacle Detection Network.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Maritime Technology

Background:

• Maritime obstacle detection is crucial for the safe navigation of autonomous surface vehicles (ASVs).
• Current advanced image-based detection methods, like WaSR, are computationally intensive and memory-demanding, hindering their deployment on embedded systems.
• There is a need for efficient and accurate maritime obstacle detection solutions for real-time applications on resource-constrained platforms.

Purpose of the Study:

• To analyze the performance bottlenecks of the state-of-the-art WaSR network for maritime obstacle detection.
• To propose an embedded-compute-ready variant, named eWaSR, by optimizing computationally intensive components.
• To evaluate the performance, speed, and deployability of eWaSR on embedded hardware.

Main Methods:

• Analysis of the WaSR network architecture to identify computationally expensive modules.
• Redesigning key network stages using lightweight transformer-based approaches for improved efficiency.
• Implementation and testing of the proposed eWaSR network on both standard GPUs and embedded systems (OAK-D).

Main Results:

• eWaSR achieves detection performance comparable to WaSR, with only a 0.52% drop in F1 score.
• eWaSR demonstrates a significant speed improvement, running 10x faster than WaSR on a standard GPU (115 FPS vs. 11 FPS).
• eWaSR successfully runs on the OAK-D embedded sensor at 5.5 FPS, whereas WaSR cannot run due to memory limitations.

Conclusions:

• eWaSR is the first practical maritime obstacle detection network designed for embedded systems.
• The proposed network offers a balance between high detection accuracy and computational efficiency for ASV navigation.
• Public availability of eWaSR source code and models facilitates further research and development in embedded maritime robotics.