Computer vision, collision avoidance

[markxengineerin]

After crashing into a piling, then 1 year later, a channel marker, I'm trying to start a security camera based "obstacle detector". I chose this camera:
https://www.hikvision.com/en/product...2CD2087G2H-LIU
for its low cost and great low-light performance. I have not tested outside yet, waiting on a longer ethernet cable, but will post when I get some images.

For the hardware and code, I'm planning to use a friend's custom hardware, which is very similar to an Orange Pi 3, and OpenCV/Python.

An equivalent commercial product is Sea.ai but what I have in mind would be cheap and open source.

I'm just wrapping my head around the different image detection models (Yolo seems most popular) and the differences between models, frameworks, training tools, etc. So far, I have a feeling the work is already done and open source, somewhere, so what remains is easy instructions for how to implement it for newbies. I hope not to need to get too deep into the weeds, want to get something working with minimal effort.

So questions, for anyone knowledgeable in the area:
1) Where to learn about state of the art, in this area? If I google "marine object detection", there are many half finished projects out there, but no way of knowing which one is the best starting point.
2) Especially interesting, are there any existing models out there that would do a good job detecting marine objects, and would not need training?
3) Assuming training will be needed, I would like to compile a list of training data sets
4) Want to help? Please get in touch...

[barnakiel]

It does not need to be vision based. If it is vision, it will be limited to good visibility and distance.


Take an imaginary trip - what if the device reads digital map data and has a built in compass. Then it can overlay map (nav lights, shore line objects, etc) + nav data (AIS objects, like boats) data on what you are seeing.


And maybe then, as a nice extra, throw in object recognition.


I have already built a prototype of the first thing.


Let me know how object recognition part goes for you. Fot it is not easy. Even Tesla still makes many mistakes.


b.