Raspberry Pi model 3B video processing speed

[chrissteffen98]

I'm coming from the Arduino world so im just starting out with Raspberry Pi. I've built an autopilot system for an rc Plane and wanted to add object avoidance to the plane. So my plan was to use an arduino for the hardware connection and then use a Raspberry Pi 3 B with a camera to process video coming from the camera and basically see things like trees and get their rough positon.
The problem is i dont know how fast the raspberry pi 3B can process video. I think ill need around 30fps but that still might change. Is this even possible?
Like i said i'm just starting with raspberry Pi and dont know how fast it is.

[blimpyway]

I doubt it. If possible it needs lots of knowledge, work and optimisations.

I mean even figuring out what is close and what not, with a single camera is quite a hard problem. With double camera (stereo vision) we know how to do it but still it's quite intensive processing quite above Pi3's ability to compute at 30fps.

I would be amazed if this could work at 10fps and be happy with it.

[amcdonley]

This is my analysis of motion detect on a “non-aspirated” Pi 3B Video Surveillance
(TL:DR 10fps heat limited)

My testing of “Recognition” such as faces, color and size, lane markings run around 1 FPS.

All these are single threaded Python3 with “stock packaged” OpenCV. One multi-processing test resulted in doubling the throughput to 2 FPS.

[blimpyway]

This is my analysis of motion detect on a “non-aspirated” Pi 3B Video Surveillance
(TL:DR 10fps heat limited)

Interesting, so the CPU loads were 1.5%-2% yet temperature was high?

PS/ Your link doesnt work well on Ubuntu/Firefox but it works fine
without quotation marks after url=

[jbeale]

simple motion detection is done on a Pi by the PiKrellCam https://billw2.github.io/pikrellcam/pikrellcam.html package basically as a side-effect of the H.264 compression, which has to calculate motion vectors anyway, so it has the same GPU, CPU and power footprint as simply recording video and it works at normal video framerates.

3D distance estimation is a much harder task and I wouldn't be surprised if 1 or 2 frames per second is where you'd end up. However you might be able to get away with working at very low resolution images, if we're talking about a mostly uncluttered environment, and if so you might do better.

[amcdonley]

Interesting, so the CPU loads were 1.5%-2% yet temperature was high?

PS/ Your link doesnt work well

Thanks. Fixed the link.

Not %. Load average is some weird number that seems to correspond to the number of cores in use when the number is less than the number of available cores. So 1.5 is an average load of 1.5 cores in use out of the 4 available on my RPi 3B.