I don't have the original link. Just search for fpv or drone and ukraine or other similar words.

custom design

custom design

Private research and demo only. No plans to share techniques. You can replicate some of these by using public opencv trackers. Code to use those pre-built libraries are easily available at opencv.

You can build a more robust tracking system with just 77 lines of code. You wont be having a Nvidia RTX 30 series strapped to your rocket or your drone. The delay between tracking and control will just be too huge for it to work. Those systems have to run in real time

A few opencv trackers are fast (like mosse), most are slow-ish. Those do not need GPU. Use depends on processing speed and flight control response. High speed stuff like rockets need specialized solutions. Drones can work with slower tracking. This demo has no control, just tracking. No GPU, no specialized library. If I can make it work at 30-60 fps, I'm happy. Good enough for slow drones.

@@timelapsecoder I tested my tracker on a cheap 2010 camera module, to be honest I think it can handle speeds of the Spike NLOS and the lock on is less jittery even with a lot of noise and shake from the cheap camera. The best solution is to do the tracking and data processing from the IMU in the same computer and feed the microcontroller the control commands. The only problem is the delay between tracking, IMU data processing and control, I dont think the link delay between the 2 computers is the biggest problem since servos have wayy more noticeable delays, i think just the processing in the first computer needs to be real time enough. Maybe between each frame, if you fed kalman filter predictions to be used as real data, maybe it would lead the system closer to where it wants to be until the real data comes and corrects it. I really dont know how well that would work, just a hypothesis until proven otherwise

If the projectile is already axis-aligned to the target and the target is not moving, then tracking workload is relatively easy even at high speeds, even if the tracker is slow. Off-bore tracking is where targets will whizz by. So as long as you aim your tracker dead center before releasing, it has a good chance of keeping track. Your ideas are correct. If the tracking and overall system response, including mechanicals, are slow, then you implicitly include those into your guidance inputs (e.g., Kalman predictions) and fine-tuning to anticipate system behavior.

I have not done that comparison.

look for mouse in opencv. drawing stuff is part of opencv.

@@timelapsecoder yup figured that out but how do i track a specific part like you do

i select a part of the image. then a tracker follows that part. you can use any tracker.

​@@timelapsecoder i meant that you are not tracking the object rather pixels so how do you do that

old style trackers track images, not objects. their logic are described in books, internet.

I made a nav system like that before, but not using gmap tiles, just drawing a route over an existing map. That required many image comparisons to match the pre-made route. In this example, the idea is to use a live feed of the target and the drone plots it own route.

@@timelapsecoder ok good! i like your videos and i work in the same field, i hope one day u will write an article! have a nice day

People acting like they have built some DARPA type stuff. I have built a more robust tracking system with just 77 lines of code. People dont take into account that those systems have to run in real time in a resource constrained environment. In his script, the delay between tracking and control to keep the target in the center is huge and for a system thats traveling close to the speed of sound like SPIKE NLOS, that wouldnt work very well

You can reach out to the poster to sell your 77-line tracker. If it works across many scenarios and fast enough while being impervious to clutter, that's good. If it runs on low-power CPUs, even better. Btw, this is demo is just a tracking overlay on a fixed video. There is no control loop delay to observe.

@@timelapsecoder Well, I wasnt trying to sell it, I thought he was talking if you could somehow sell the tracker to the gov or something else. To sell a tracker it has at least to be written in C++, be robust against large occlusions and wide range camera movements

@@orion_stelee For unpredictable environments, not yet good enough. For simple stuff like line-of-sight follow-me drone feature, this will work.

@@orion_stelee I am also interested in understanding this topic. Can I get in touch with you to briefly find out where I should start and how to achieve the effective tracking you describe?

syempre! kumustra ka dyan?

look for line following robots. same principles.

Random videos from the internet or youtube. They are easy to search.

The video tracking part can fail under certain conditions. And you also have to add the flight control code based on the tracking.

@@timelapsecoder how can we make it better? Is there any library to train the cnn on drones? I am thinking of making a drone hunter

@@3alabo not aware of a drone dataset. this is not cnn object detection, just old style tracking. it tracks whatever it is pointed.

As long as the output is digitized, e.g., cheap USB cameras. Analog FPV can be captured into a remote USB RX and read as a regular PC camera.

I don't have it, but it is easy to find, or at least something similar. Just search for "Saab Viggen" or just viggen. Maybe add "airshow" or variations.

Thanks. Not exactly same. I have not updated the version on the video. But it is built off the same original code, just with many small changes here and there.

Not fixed. Just re-specified the target.

@@timelapsecoder Of course. I also have this. The algorithm works by detecting contours. And this is his work. But these algorithms have many advantages. They hold the target well at any image quality, if there are no obstacles (mentioned above).

My code is not looking for contours but my algorithm is probably automatically finding contours as a side effect? I find contours fail against intense background clutter.

@@timelapsecoder I used standard tracking algorithms and noticed that they probably use (maybe not only) contours. And also have such a side effect.

@@OleksandrYushchuk8 привіт, зараз пишу систему доведення для fpv, не міг би підказати, чи розібрався ти з цією темою? Найбільше цікавить, який алгоритм краще всього підходить для відслідковування цілей при поганій якості зображення та за умови, що противник бігає по по посадці в камуфляжній формі.

Solving FPS depends on your exact case. Maybe try just one model instead of many or use a smaller video? You can also try to look for models that trade accuracy for speed. My methods have no architecture, just a bunch of specialized CV logic.

Thanks. Just private code. Ground station is great for better tracking computations, but video transmission might be poor sometimes.

@@timelapsecoder Yeah the video transmission is the only issue im worried about. Thinking of using the Eachine ROTG02 so I can just use anything with a USB port for a video display.

Well that and powering two small magnetic actuators but I might be able to use the extra esc pads or change some flight controller pads around.

Learn opencv and image processing/computer vision techniques. You can choose to only read enough to test your ideas. Easier with PC then moving to other platforms later, but use whatever hardware you have now.

Thanks. No plans, just a side hobby for now.

@@timelapsecoder I bet Brave1 would support it if you took it further

custom made

Don't have the link. You can search for airshows to find something similar for testing.

you write that it is open cv

private research

Thanks 😁

Stuff like this has many peacetime uses. But in KZfaq, people click things for combat. "Tracking a ship," not a thing. "Anti-ship," that's good. :)

@@timelapsecoderlol good point. I’ve been following both your channels for awhile now and am super impressed as always, but I’ve always wondered something. Was this all self taught? What’s the timeline for completing something like this with only 1 person working on it? How do you work a 9-5 and still find the time? I dropped out before learning any math higher than linear algebra and now weld for a living, but the stuff you do is what I’ve always wanted to do. Creating your own HUD with tracking and different targeting models is so cool man. I want to learn embedded so I can create my own drones eventually. Thank you for the inspiration.

Thanks. Not self-taught, but this can be self-learned by reading about CV techniques and just trying out things. This takes time because I don't work on it all the time. I do it on evenings or parts of weekends whenever there is nothing else to do. You don't need to study a lot to do something similar. Opencv has ready-made tracking solutions. For opencv, you need basic python. Custom code is hard, re-using pre-made libraries is easier for beginners. The HUD is just lines and text you put over a video stream. Good luck!

You can't afford me bro I'm joking You can find places in the internet where drone experts meet You can ask there You will easily get many who can help you Much better than me

@@timelapsecoder Thanks bro

Not possible. This is just private research.

in the title, color detection

@@timelapsecoder yes , i tried color detection use HSV but i cannot get the same results

If your using an OpenCV tracker that is heavily dependent on template matching, then you need to increase or decrease your template. You can write code to estimate how fast you grow or shrink. This can be sophisticated or simple, depending on your application. Simple is hardcode math logic. Sophisticated is you use information from the video to estimate.

I don't do it so I don't have an answer. Continue studying CV. Some trackers can do that.

@@timelapsecoder thx

No, but some drone websites offer tutorials.

No initializations, other than maybe basic ones like approximate size of target, but even this is not that important. The algorithm has to immediately figure out what it needs. This is the 'magic sauce'. :) It does not work correctly all the time, so not so magic.

For that you need detection and identification, which requires more computing power, and prioritization rules. The first part is hard, but the technology exists --KZfaqrs with limited AI knowledge make (copy) stuff that identify cars and pedestrians. But professional militaries do not use full autonomy. They always have a human in the decision loop, for good reasons. Weapons that auto-select and auto-attack are very controversial, as they should be.

@@timelapsecoder Yes, I have already ordered all the components that I should need to build a drone that follows humans. It was interesting to see if it was possible to create drones that were like "loitering munitions". In the sense that after having turned them on and put them in flight they "turn" looking for a target

That's how military drones work. But they are quite sophisticated, large, expensive and not for civilians. If you want military use, just reach out to a defense drone manufacturer. If you just want to create a survey drone for peacetime use, buy from a commercial drone company. Many of them would have what you need or can customize one for you. That's what they sell. Much easier than building one yourself.

Thanks!

People like to watch, and The Algorithm promotes, videos that are like this. If I track something harmless like a balloon, no one will click. Maybe I'll track a cat next.