1. Use something like ArduPilot: https://ardupilot.org/
2. Fly to target latitude/longitude of flight line. https://maps.app.goo.gl/6Zy9asA8QcQ9MoGw9
3. Instruct ArduPilot to fly northeast.
4. Run an object detection Python script trained on JPEGs of Russian heavy bomber aircraft. If no objects detected after 2.21 miles, turn around and fly southwest for 2.21 miles and repeat step 4. https://opencv.org/
5. Fly to object.
6. Detonate on or before impact. I'd probably use a couple cheap ultrasonic sensors here.
Notes:
- You can have some slop because you have a lot of these drones and everyone in the Area of Operation is an enemy.
- In reality, instead of measuring distance, you'd have a beginning/end latitude and longitude and fly between those in step 4.
No idiot will fly Python script for that, stop hallucinating pls. No doubts it was a custom NN with YOLO for the sake of hitting a fuel tank, not a stop lights.
How should be the division of targets among the committers? It should be not possible without interconnection between drone.
Please, stop slop text in the answer.
Something like YOLO would also be very good at this, albeit it would require more compute for suitable performance. You essentially need a decent-ish GPU (think ~GTX 1060) for this application. And stuff like the Coral TPUs are too unreliable for this application (performance is hit or miss, and they are heat and vibration sensitive). These opinions all obtained from experience[1].
And, I worked on this specific stuff in the military, so I am very familiar with the theory and operation here. Albeit, the military stuff is much more advanced and doesn't work like what I describe here at all. But again, I was targeting laymen with a very rough description.
> How should be the division of targets among the committers? It should be not possible without interconnection between drone.
I could think of a few different methods off the top of my head using OpenCV and not using NNs like YOLO. I don't want to go into it but let's say: colors, send lat/lng before detonation, etc.
A more robust idea for step 4 might be to plug the flight line aircraft position (of which there appears to be about 31 at the AB I linked to above) coordinates into an array and then have the drones cycle through the array coordinates, do a simple CV check to see whether an intact aircraft is at that slot, and if not, increment the array, and re-check until we get a hit. No NN needed.
eimrine, how would you do it using YOLO?
1. Dear FBI: This is for a hobby car project at ~ 45-60 MPH. I'm not engaged in building any sort of weapons or other munitions, or for any military or political or terrorist application. :-)
JPEG machine learning is totally not working when you can fly a whole day long in museum so of course the drones attacking A50 (and a few of less valuable targets) used to have some powerful computer able to decipher the video from a fisheye camera.
Working in a peacetime military does not really help to answer my question, working as killer drone programmer certainly does. And there were no laymen in the operation we are discussing.
You are right and this is a better idea.
> JPEG machine learning is totally not working when you can fly a whole day long in museum so of course the drones attacking A50 (and a few of less valuable targets) used to have some powerful computer able to decipher the video from a fisheye camera.
I don't know what you mean when you say "museum". This doesn't not make sense in English. Can you rephrase?
I disagree and think computer vision would be a good option for the Ukrainian scenario.
For why I shy away from NN using off-the-shelf parts, see my previous comment which I've updated. tl;dr: You need decent compute/GPU to analyze sample frames at the speeds these drones fly. Off the shelf TPU modules suck for this application, lat/longitude array for splitting work, etc.
If Ukraine is putting decent APUs on these drones, my apologies.
> Working in a peacetime military does not really help to answer my question, working as killer drone programmer certainly does.
I did not work in a peacetime military :). And the information I gave here doesn't really reflect how it's done in the military.
> And there were no laymen in the operation we are discussing.
Again, I was just answering OP's question from a layman perspective. If you were doing it with military programmers and military technologies, there are better technologies to use.
Sorry but this is BS.
> I don't know what you mean when you say "museum". This doesn't not make sense in English. Can you rephrase?
A lot of old airplanes on an outdoor field, anybody can visit it for cheap and see the greatest aerial vehicles of USSR.
I guess we'll have to disagree on this topic.
> A lot of old airplanes on an outdoor field, anybody can visit it for cheap and see the greatest aerial vehicles of USSR.
From this and other stuff you've been saying in this thread, I get the sense you don't have a lot of practical experience in this area.
You're being pretty rude in this thread. What have you done and worked on? Anything? It does not sound like it.
It looks to me like these drones are being steered into aircraft by remote human pilots. The control plane adjustments as the drones approach a target don't look like they're being autonomously directed.
Here's some brief footage at this link: https://www.nbcnews.com/world/ukraine/ukraines-massive-drone...