Testing Freds new program on the Nixie build 1.1.429.

 

 

Good news is it's working :-)

While I was in there, I started OpenCV, loaded the PyramidDown Filter and the Flip filter.

I have mounted the camera in the eye ball upside down (Doh!)

then I started YOLO :-)

It's Slow, but it is worky :-)

 

kwatters

3 years 6 months ago

Awesome! Yeah, yolo is very slow. But in a way it's amazing that it's worky on the pi.

It was a multi project effort to make that happen. Some respect needs to be paid to the javacv open source project to make that a reality. In a way they have been front running things that we've wanted to do in mrl.

It's putting the bounding box where it saw the person in your video... At that point there are a bunch of frames waiting to be processed. One area that I hope we can improve is the latency of the video through the video pipeline.

I did some work on the mjpeg frame grabber to configure it to drop frames when it is falling behind. But, I guess, for now. Horray for worky!

All things considered, the frame rate on the raspberry pi4 with all the other services I had running at the same time, is actually very good.
between 20 and 30 FPS.

Much better than the early 5 FPS with nothing else running.

What I observed is You grabs a single frame and then processes it for the next 90 seconds,
then it will output the results with the bounding box and grab the next frame from the live stream.
90 seconds or so later it will output the results again continuing the loop.

If I put a much more powerful computer there, it would of course be faster. but it's still pretty amazing at how well it works, especially considering I had the rest of Fred's processes and services running at the same time on the same Raspberry Pi.