What up anons,What do you guys think of my group project for undergrad university.Creating an AI model that identifies military aircraft and identifies the direction and velocity they are flying in relation to a camera. How should I start about developing this (we have 3 people in our group project, and 2 months to complete). How would you architecture the project?When we finish the project we plan to publish it on github with MIT license.
>>107070250Maybe it might be better to use Affero General Public License so when we this model is used to serve against clients (planes) they can get the source code.
>>107070250Contact Palantir.
>>107070250You don't need "create" anything you can just train a dataset for YOLO or whatever.
>>107070250not doing your homework, ranjeet
>>107070250First, that's not just one AI model.How I'd do that, is start training classifiers, binary if you are noobs, multiple if you're any good. The classifier will be responsible for determining the type of aircraft, ideally, you'd need a TON of pics from different angles, configs, etc. of the bogey.Secondly, I'd use IFF, so I'd really train on sqn/flight, or at least country markings.Second: inferring the current heading is easy, you'd need to identify the beginning and the ending of the nose, where the nose points is where the plane is going. I'd do that with simple image manipulation. For extra credits, I'd start looking at flight surface and exhaust jet orientation (typically Russki jets has thrust vectoring), that would give you clues about the future heading of the bogey. However images you usually get from intel are often too shit quality to make the flight surfaces clearly visible, especially since small deflections can cause a large change in bogey attitude, and bogey can course correct in < 1seconds.Third:For velocity, you need two images, preferably knowing the difference in camera location, and image timestamp.Use your database to infer the distance from camera of the plane (you know how large a plane is in real life, how large it is on camera), and then try to calculate how many pixels it has moved. Draw the curve of motion (planes can move along splines only), estimate the length of the spline, then use the timestamp difference.If you know the absolute location of the cam, you can estimate bogey location, and altitude as well.So your system is:> Train models to classify planes> Cut image to plane size> Use plane sized image to identify attitude> Reverse ray trace to identify plane distance from camera> Classify if plane is still the same on second image> Create the movement spline of plane> v = s/t.
