Depth performace with drone

RyanLee

Hi,

We tested OAK-D Pro FF (with ir cutfilter) on our drone.

Here is the test video.
test video

And stil image.

And i got some issue with cloud point. Such as there is not object but the cloud point is there like below.

Could you guide us to enhance it??

Best regards,

RyanLee

and for this test. We used below repo.

<source>
https://github.com/njezersek/depthai-experiments/tree/rgbd-point-cloud/gen2-pointcloud/rgbd-pointcloud

<point you giude us before for the drone company>

stereo.initialConfig.setMedianFilter(dai.StereoDepthProperties.MedianFilter.KERNEL_7x7)
config = stereo.initialConfig.get()
config.postProcessing.speckleFilter.enable = False
config.postProcessing.speckleFilter.speckleRange = 50
config.postProcessing.temporalFilter.enable = True
config.postProcessing.spatialFilter.enable = True
config.postProcessing.spatialFilter.holeFillingRadius = 2
config.postProcessing.spatialFilter.numIterations = 1
config.postProcessing.thresholdFilter.minRange = 400
config.postProcessing.thresholdFilter.maxRange = 200000
config.postProcessing.decimationFilter.decimationFactor = 1
stereo.initialConfig.set(config)

erik

Hi RyanLee ,
I'd suggest maxRange threshold filter, and pointcloud postprocessing where you remove outliers.
Thanks, Erik

RyanLee

HI erik

Thank you for your feedback. we already tested that maxRange threshold filter. But we thought that it is temporary solution not ultimate solution. and above issue is coming within 8m. if you have any more idea about it please let us know.

Best regards,
Ryan.

erik

Hi RyanLee ,
Are these points from the corner of the depth image? One option that we have seen users to go with is to crop the depth eg. 10pixels around the image, where depth can be much noisier, and that helps with pointclouds.
Thanks, Erik

RyanLee

Hi erik

My concern in the previous question was in the pointcloud of yellow circle. Because there was no object but we can see the pointcloud. And it happnes with in 8m.

Here is the image for your information. And i also add the comments in the image around yellow circle.

Best regards,
Ryan.

erik

Hi RyanLee ,
Regarding image above - could you elaborate on what you have shared? Have you also tried voxalization and removing statistical outliers?
Thanks, Erik

RyanLee

Hi erik

Ok i will explain about it once more.
the device for this test was OAK-D-Pro (with ircutfilter). And our application is drone. And our test was outdoor environment. (F.Y.I, indoor test result was so satisfied so we want out outdoor.)

But we met some issue. In case of within 8m test, there is some ghost pointcloud is detected like above images with red circle. Again the ghost pointcloud means that when our drone is flying, even though there is no object in front of our drone, we can see the ghost pointcloud. So it makes some issue now because based on the characteristics of drone if the drone meet any object, the drone should stop to avoid any crash.

Again, the ghost pointclould made our drone stop even though there is no object in front of drone in the sky.

And we tried all the filter for a long time but it made another side effects. So using filter is not a solution for this issue based on our verification.

If you need any more information, feel free to tell me.

Thanks.

erik

Hi RyanLee ,
Could you provide left/right images as well that produced this depth/pointcloud?
Thanks, Erik