How exactly does obstacle avoidance work?

[PaulJohn]

Hey,

First of all I'd like to say I love my P4! Best bird ever!

I'd like to gain a better understanding of how the front and back cameras perceive the world around the drone, and how it calculates the distances between obstacles. An in depth explanation of how the entire system works would be ideal. Some pictures would hurt either. Do you know where I could find such info, besides DJI's promotional videos which don't really shed much light on the subject?

Thanks,
Paul


Sent from my iPhone using Tapatalk

 

[bradbrok]

I found this, basically it's a very in depth subject matter that's still being developed.

The word you're looking for is "Stereo Computer Vision" which is able to extract depth information from 2 (or sometimes more) cameras. It compares the same object from two views and calculates the approximate difference between both scenes. The more difference there is on an object, the closer it is.

http://www.cs.unc.edu/~lazebnik/spring09/lec14_multiview_stereo.pdf

That said, if an image is not clearly defined or has no recognizable pattern, it'll be impossible for the P4 sensors to pick up, I do however suspect that the OA sensors actually can pick up IR light as it may have no IR filter on the sensors themselves. But my testing shows that it definitely picks up pretty small objects. Long skinny objects are a bit more difficult for it.

 

[mactech]

There are no backward facing cameras.

 

[bradbrok]

There are no backward facing cameras.

I think he's talking about the downwards facing VPS that works similarly to the OA system. Same concept applies to both.


Sent from my iPhone using PhantomPilots mobile app

 

[Shammyh]

Some sort of Photogrammetry?

Sort of how your eye/brain uses binocular cues to infer depth (I can catch a baseball, but I don't have a laser distance sensor integrated into my brain). The Phantom is running some relatively basic algorithms to accomplish the same thing. The cameras are pretty low res, so I'm sure it's fairly coarse measurements, but more than enough for basic obstacle avoidance.

Basically as others have said, it's comparing the two images from each camera, inferring a measure of depth/distance, and building a model out of that data.

Truthfully, it's not really new tech . But the miniaturization and integration into the flight controller is pretty novel.

Hard to say exactly what algorithms they're using, as it's not published information, but I doubt it's very different from other solutions out there in cars/robots/etc.

 

[Shammyh]

For a very simple example of how this concept works, you could also look at: Parallax - Wikipedia, the free encyclopedia

Modern photogrammetry is significantly more complex, but the basic concepts are similar.

 

[PaulJohn]

Nice, thank you! What a captivating subject!


Sent from my iPhone using Tapatalk