Computer Vision

RE: Computer Vision

boshay — Wed, 19 Aug 2020 23:33:29 GMT

You can use a regular USB web cam assuming your drone can handle the added weight and the software is OpenCV a free computer vision library that you can use to get what you need up and running with a little python script. Also you can find a bundle of tutorials online for OpenCV and python. I have, myself, tested a simple object tracking app with OpenCV on the Raspberry pi 3 and know that it works perfect.

RE: Computer Vision

aswinvenu — Tue, 21 Jul 2020 03:20:32 GMT

As Doug was saying it's an interesting and ambitious project. What is the rough area you will scan at a time? Bear in mind that aerial scans not always reveal you the number plates because of the angle. So your drone should be scanning always at an angle. Do you want to include moving objects for scanning ? Like a cruising bus or a car ? Then you may need a high speed camera. Normal cameras capture videos anywhere from 30-50FPS. That's not sufficient to capture moving objects without blurring effects.

You have mentioned that the image recognition will run on your PC remotely. How far is your PC sitting from where the drone is operating? How are you planning the drone to send images/videos to your PC?. Streaming high resolution videos/images is not reliable over long distances. So doing the image recognition on the drone is more reliable I would say. But you need to master the algorithm before anything. Train it well and deploy the inference on the onboard computer (May be it can be Rpi4 but consider TPU accelerators like Corel https://aiyprojects.withgoogle.com/edge-tpu/ ) The more light weight your inference program the faster it process the frames there by higher frames per second resolution.

Things to consider

wide angle and high resolution cameras - More area covered in one frame and zoom to detect number plates without loosing the image quality - reduced flight time
Image stabilizer / Gimbal for capturing stable images
High FPS Camera to capture moving objects at faster speed
Scan at 45 degree or low to get the number plates more visible ( else all the letters appears like ellipses and leads to false positives and false negatives)
Add intelligence to the scanning. If you identify a vehicle and can't get the number plate then fly down and get a better view of the vehicle and search for number plate. Still its not visible then black list the vehicle and store the images.
Use AI inference accelerators like Google's edge TPU to speed-up the inference
Tight power management

Regards,

Aswin

RE: Computer Vision

dougw — Tue, 21 Jul 2020 01:54:50 GMT

This is a very ambitious multidisciplinary project with lots of potential applications. However if you want to complete it in a reasonable time, you will likely need to buy everything you possibly can and only work on the bits that can't be bought or found open source. As in buy a good drone with good camera. Find the best vision recognition software you can and buy a big computer to run it on. Use a separate computer and software to interpret your commands and generate intelligent search pattern control signals. Stream video from the drone to your big computer for processing. It will still take a lot of effort to get these systems working in the way you want.

A Raspberry Pi4 is a great little computer, but asking it to do real time video analysis with commercial or open source software is asking a bit much. If you are a wizard programmer who is also a video analysis guru, maybe it could do something surprising, but it would not be easy. I really don't want to sound negative, I like the idea, and think it would be great if you could make it happen, I just find the scope of the project to be scary big.