Building an Open Maritime SAR Perception Stack on RK3588 Edge Hardware

Modern AI models can achieve impressive results in research environments, but deploying those same models on real hardware is often where projects begin to fall apart.

While working on UAV and computer vision systems, I repeatedly encountered the same problem: powerful models running on expensive hardware in controlled environments, while practical deployments were constrained by cost, power consumption, compute limitations, and the realities of operating in the field. On one end were systems that required hardware budgets difficult to justify at scale, on the other were low-cost platforms that often struggled to deliver reliable real-time inference. The gap between "it works on my GPU" and "it works on a drone" turned out to be much larger than expected.

SentinelBlue started as an attempt to bridge that gap. It is an open-source project exploring how far practical computer vision can be pushed on affordable edge hardware while using open-source tools.

This talk follows the journey of building an open maritime search-and-rescue perception system. I will discuss how maritime SAR datasets were curated and balanced, why multiple YOLO architectures were benchmarked instead of blindly selecting the latest model, and how deployment constraints ultimately influenced design decisions far more than model architecture alone.

A major focus of the session will be the path from training to deployment: exporting models from PyTorch to ONNX, navigating RKNN compilation requirements, dealing with quantization trade-offs, and targeting RK3588-based hardware such as the Radxa ROCK 5C. Along the way, I'll share lessons learned, what worked, what failed, and which assumptions break down once you leave the desktop GPU behind.

Rather than presenting AI as a cloud-only technology, in this session we will explore how open-source software and accessible hardware can be combined to build deployable perception systems for robotics, UAVs, and embedded applications, where I hope to leave the attendees with a realistic understanding of the challenges involved in taking a computer vision project from a notebook to a device that can actually run it.

The talk is intended for developers, makers, robotics enthusiasts, students, and embedded systems practitioners interested in edge AI, open hardware, computer vision, and real-world deployment workflows.

• Understanding why many computer vision projects fail after training: the gap between "it works on my GPU" and "it works on a drone", including real-world constraints such as power, thermals, memory, cost, and deployment environments.

• A practical roadmap for deploying AI on open hardware, covering the complete journey from dataset creation and model training to ONNX export, RKNN compilation, INT8 quantization, and embedded execution on RK3588-based platforms.

• How to choose the right hardware for edge AI workloads, including lessons learned while evaluating Raspberry Pi, Jetson, and RK3588-class devices for UAV and robotics applications.

• A deep dive into the PyTorch → ONNX → RKNN deployment pipeline, including operator compatibility issues, static graph requirements, conversion failures, quantization challenges, and embedded optimization techniques.

• What actually matters in production computer vision systems: balancing accuracy, latency, memory usage, hardware constraints, dataset quality, and deployment feasibility instead of optimizing purely for benchmark numbers.

• Practical engineering lessons, failed experiments, design trade-offs, and open-source tools that transformed SentinelBlue from a machine learning project into a deployable UAV perception system.