Jetson Nano support

[anywaterdev1000]

Hey,
We are planning to build a trimaran USV as a team. We are thinking of using 2 cams. We are thought that using Jetson Nano Dev Kit would be good bcs there is so much data to process (it has an external gpu included so its powerful for image processing). Any chance for compatibility using an Ubuntu OS?
Regards,
Salih

[seandepagnier]

I am also looking at the jetson platform, but was hoping to get the xavier but it is just out of stock everywhere.

pypilot should run on ubuntu platform. As long as the gpio works and i2c and uart work, then pypilot should be supported. I cannot confirm it yet but if I ever get a jetson nano I could try it.

The next versions of pypilot will support image processing with cameras to enhance performance however the jetson nano is even a bit limited for video processing with training on the fly. It should be sufficient for inference, but what resolution it can process video and actually do learning is in question.

What is your plan for cameras? My thought is one camera that can see the sails and another camera that can look toward approaching waves.

[anywaterdev1000]

Good that u reminded. Even a Nano might be uncapable. We are preparing for an autonomy competition so probably a front camera to inspect obstacles and markers and the second camera might be a supportive camera to get a distance refference using a comparison method between the two images. Still Im not fully sure maybe a lidar or a radar will be a better pick in terms of distance measurement but we are a bit tight on the budget. maybe a third camera thats at the back of the ship just in case we need it.
btw the ship will be fully powered not sailed.

[seandepagnier]

Well, for me anyway, fully powered means sail power. The camera aiding in steering the way a human benefits from vision when steering a boat. Any additional sensors like lidar, water speed etc would likely give more useful data for the processing power than a camera.

I also think even a raspberry pi 4 can possibly be used (for inference or at least mostly transfer learning) but at what resolution? It may have to down sample the camera input to 32x24 or 16x12. This would not give much more data than seeing a horizon but even that is useful, but being able to learn as well over time takes much more cpu. If you do training offline it is a different problem, but for pypilot it needs to retrain and when things change like sensors or camera angle so online training is most useful and this needs a lot of calculations. Also for development more processing power is desired but I cannot find the other nvidia jetson (more powerful than nano) anywhere, as the supply chains have failed.

This has nothing to do with obstacles or markers. As for not being powerful enough who knows. For object detection I know of:
https://github.com/mavixorg/open_watchman

Let me know if you have any similar projects. It is a difficult problem. If you constrain it to a particular boat on a course or something becomes much easier but also not very useful in practical cases.