Optical frame change due to temperature?

procopiostein 2016-07-29 13:12:57 UTC #1

Hi there,
I am using an Astra camera in my robot and the camera is tilted, facing the floor, and I noticed that during usage, I have an effect as if the camera was changing its orientation (roll).

To calibrate its position, we use an algorithm which detects the main plan and calculate the sensor height, roll and pitch.
I have noticed for a while that I constantly needed to recalibrate my cameras and decided to investigate it further.
What I've found out is that there is some kind of error that grows with time and eventually stabilizes, as if the camera was changing its roll (which it is not, it is mounted in a very stable setup). I took measurement for one hour, starting after 30 minutes with the camera off, and this is how it looks like:

time - computed roll (degrees):
13:53 - 0.88
13:59 - 1.27
14:06 - 1.76
14:12 - 2.10
14:17 - 2.25
14:32 - 2.33
14:38 - 2.37
14:52 - 2.40
14:57 - 2.41
15:09 - 2.40

do you have any idea what is going on? I am guessing temperature...
can you think about any way to overcome this situation? maybe the temperature compensation is not working as it should...
any ideas are more than welcome.
thanks.

qlkvg 2016-07-29 14:00:05 UTC #2

I have the same effect, both with astra pro and astra s. fortunatly, i do not need depth sensor to work continuously, so I stopping it after some timeout and enabling it, when it needed. this is the only solution, that I found.

procopiostein 2016-08-01 16:38:21 UTC #3

thanks for your answer, @qlkvg. do you have any hint on what causes this effect?
if it happens to you with both the astra pro and astra s, and for me with at least two astra, it seems to affect all the astra models.

rok 2017-02-02 16:06:11 UTC #4

Hi,
Any news on this? I'm evaluating possible substitutes for Asus Xtion Pro Live for my applications, but need a camera with a stable behavior, since it has to work for an entire day continuously..

procopiostein 2017-02-02 16:18:58 UTC #5

well, the news is that this issue has been confirmed by Orbbec, but so far no fixes for it....

zhumxcq 2017-02-13 15:02:00 UTC #6

I think Xtion has this problem too, so does Kinect. It has to do with thermal expansion of the laser diffraction grating.

Vale68 2017-05-31 09:33:51 UTC #7

Same problem with at least 3 astra, any news about that ?

procopiostein 2017-06-01 06:18:02 UTC #8

unfortunately not. It seems the orbbec team is mostly concentraded on the Persee nowadays =(.
does anyone with a Persee can confirm it is also affected?

sarah 2017-06-01 09:08:14 UTC #9

The problem exists on the Persee too.
The persee is static.
We select some points on the floor to calculate the floor, the height.
After some time, it is as if the persee calibrates itself and the floor does not appear straight.
I don't know if this is a temperature issue but this is a big problem.

procopiostein 2017-06-01 09:49:33 UTC #10

Thanks for your feedback @sarah. I agree, depending on the application this is a HUGE problem. We had to ditch the Astras because detecting the floor is crucial for our application and the data becomes useless after a couple of hours.

It seems now that all the Orbbec RGBD cameras are affected by this issue, Astra, Astra S, Persee... this is a big issue and I hope they address it sooner than later.

@agnes, @josh, do you have any info on this matter?

Vale68 2017-06-01 10:09:20 UTC #11

@procopiostein: you ditched the Astra for what ? I mean does exist a similar product without this problem ?
Thank you

AReimann 2017-06-29 05:34:34 UTC #12

I made a ros package for on the fly calibration using the ground floor plane. It works well enough for our robot to negate the effects reported here and even makes the extrinsic calibration partly obsolete.

The documentation (= readme) is probably not very approachable yet though.

procopiostein 2017-06-29 06:11:47 UTC #13

wow, that is great @AReimann! I'll try it asap, thanks for sharing!

procopiostein 2017-06-29 06:13:22 UTC #14

@Vale68, we are still trying to figure out the best replacement. We tried the Intel R200 but the range is quite short and we had trouble detecting dark surfaces.

victorcp 2017-10-05 08:41:09 UTC #15

Any new on this issue?

We are experiencing the same problem with the astra mini camera.

We have tried to use the package proposed by @AReimann, but we could not make it work for the Ros-Kinetic distribution.

Thank you.

v-lopez 2018-05-17 07:30:19 UTC #16

Still happening with Orbbec PRO cameras.

You can see the difference after 3 hours of use. There are 2 aruco markers on the main plane for reference.
To us it looks like a 20º difference.

Unfortunately we cannot use the package by @ARelmann because we're not facing a plane the whole time, this was supposed to be a sensor for manipulation.
We'll have to ditch these cameras, any good alternatives that have RGB and microphones?

v-lopez 2018-05-17 07:30:51 UTC #17

Here's the original situation for reference:

ZhivaevDmitry 2020-02-07 20:11:54 UTC #18

Impact of Thermal and Environmental Conditions on the Kinect Sensor