Please look at this video from the minute 11:40 and let me know where the APAS was while the drone was flying against the pole.
What went wrong?
Was the white color not recognized by the algorithm? Didn't it detect the edges of the obstacle?

It was flying in cine mode..so very slow.
And a few minutes ago he had recognized the trees and branches.

... I'm horrified too!

Overlook such a big obstacle .....

However, with this video we can help DJI to verify the problem. They can do it themselves and see what happens.

I predict we're going to see a lot of these collisions. While you'd hope that the rusty texture on that goal post crossbar would be trackable, the bar is largely featureless. When the two cameras see the same horizontal feature, the stereo vision depth perception can't obtain a fix. We're going to see collisions with power and telephone cables. It's a certainty.

In order to prevent that, you would need three cameras per orthogonal direction.

This means that pilots will still have to pay attention to their surroundings (shocker!) especially looking out for long horizontal features. The coverage is significantly improved, but it still has week spots, Look at the location of the APAS cameras and imagine each pair being your eyes, each pair will have difficulty calculating the distance to a line that "joins" the two.

Hope that helps

Nicodema Posted at 11-18 06:41
I predict we're going to see a lot of these collisions. While you'd hope that the rusty texture on that goal post crossbar would be trackable, the bar is largely featureless. When the two cameras see the same horizontal feature, the stereo vision depth perception can't obtain a fix. We're going to see collisions with power and telephone cables. It's a certainty.

In order to prevent that, you would need three cameras per orthogonal direction.

Your thoughts are very impressive, really! Thanks.

But of course these are not good prospects .....

There can be two things:

the system did not calculate the distance well
the system did not see the bar at all.

I don't know if they do an edge recognition before making the depth estimate.

The system should have detected this considering the size of the object within the scene. This technique is also usually used to determine dimensions and distances

I think its monochrome. Its pure white and this type of sensors cant't be seen by this type of sensors

It's not the colour, but the long horizontal shape that's the problem. Imagine standing in front of a very taught, horizontal clothes line and walking slowly towards it. Without tilting your head to one side, how would you tell how far you were from the clothes line? You couldn't until you were so close that you could make out fine detail on it.

The crossbar was very much like the clothes line for this drone.

In my humble opinion, they should have left the oa out of their drones.
It is very false security.
That money and weight could have gone into either a lighter drone, bigger battery or better camera system if you ask me.

I tested the same situation with my Mavic Air2 and it stopped! I also tested with a thin cable (3 telephonic cables twisted) and it stopped!
So it means that the problem is not the horizontal bar but something inside the firmware or something around the new use of the fisheye cameras.

On the air2, however, I noticed that it sees the obstacle better if it is slightly below its position. I'm talking about a few cm. Not flying right in front.
Obviously, if the horizontal bar is at least 10cm wide, it can be seen very well.

My tests were carried out with horizontal bars of different colors and different sizes and with different positions with respect to the background.

Raf_IZ0QWM Posted at 11-19 22:53
I tested the same situation with my Mavic Air2 and it stopped! I also tested with a thin cable (3 telephonic cables twisted) and it stopped!
So it means that the problem is not the horizontal bar but something inside the firmware or something around the new use of the fisheye cameras.

Very well researched, helpful!

The amount of video data (all around) is also huge ...
Perhaps it is also helpful if the data arriving in the direction of flight are analyzed with "higher priority" ...
But DJI has certainly gained a lot of experience on this, and hopefully the firmware will be upgraded accordingly.
The case of a crash presented by you is really unforgivable, is absolutely up to DJI !!!

It is good if such tests are carried out by the users (in a manageable environment), but "in real life" you should actually always fly as if there were no obstacle detection!

Raf_IZ0QWM Posted at 11-19 22:53
I tested the same situation with my Mavic Air2 and it stopped! I also tested with a thin cable (3 telephonic cables twisted) and it stopped!
So it means that the problem is not the horizontal bar but something inside the firmware or something around the new use of the fisheye cameras.

We’ll it’s good to hear that DJI’s algorithms in general can spot horizontal features. I imagine that the more texture they have the better, since I think my assertion holds that if I couldn’t use my stereoscopic vision to tell how far a horizontal wire or bar was, then a machine would also have trouble. Once there is texture or writing or other features on the “bar” then it becomes easier. You would have hoped the rust on that goal crossbar would have helped.

I think your point about the fisheyes lenses might be significant, as any narrow feature will occupy less of the frame than before. I’d hope they can tune the algorithms over time.