Fluctuating and Unstable depth stream from Intel real sense D435.

[idata]

By seeing the on-paper specification, I have bought 6 Intel Real sense D435 sensors. And the quality of depth information is too bad when compared with Kinect V2 sensor. For my application, as a first step, I need to do background subtraction. While I was using Kinect V2 sensor, I used a static background depth frame as a reference for separating the background and foreground pixels in a depth image.

But with Intel real sense D435, the depth information of background (like walls, doors etc) are always fluctuating. Even I tried with different pre-sets, and postprocessing filters, but the depth frames are not stable. With the depth quality tool from Intel, the plane close to the sensor(<2m) has Plane fit RMS error <1mm. But the background walls in the room which are around 6m from the sensor or even some objects around 2m are not stable.

Attached are the videos showing background depth and point cloud form Intel real sense D435 and Kinect V2 sensors.

How can I achieve the quality of depth similar to Kinect? Are there any known settings to get depth frames without any vibrations?

By the way, for my application, I need depth frames at 60fps. So I am using 60 fps with a resolution of 848 x 480.

Thanks in advance, for your help and time.

[MartyG]

It is worth mentioning that Kinect and RealSense 400 Series cameras cannot be compared directly because they use different technologies. Kinect 1 uses Structured or Coded Light and Kinect 2 uses Time of Flight. The 400 Series cameras use Stereoscopic imaging based on left and right IR sensors.

Excellent tips for tuning and optimizing 400 Series cameras and the images that they capture can be found in the presentation document linked to below.

https://realsense.intel.com/wp-content/uploads/sites/63/BKM-For-Tuning-D435-and-D415-Cameras-Webinar_Rev3.pdf https://realsense.intel.com/wp-content/uploads/sites/63/BKM-For-Tuning-D435-and-D415-Cameras-Webinar_Rev3.pdf

Looking at your video though, the most obvious candidate for the cause of your image problems is the multiple number of what seem to be strip-lights in the ceiling. Unlike bulb-based lights, strip lights are known as fluorescent lights because they contain a gas. This gas flickers at frequencies hard to see with the human eye and can cause noise in the stream.

[idata]

Hi,

Since I am using 60 fps and these fluorescent lights flicker at 60 Hz (power supply), I don't think that is causing any problem. Even I tried with turning off the lights, the results are the same. Anyway, I will try tuning my sensor based on the provided document.

Thanks!

[WTatt]

FWIW - with the D435 you are unlikely to every achieve depth data of the quality of the Kinect 2.0 - its just not technically possible.

Read up on RMS errors in the tuning documents = the DISTANT parts of the frame are likely to fluctuate worse and worse depending on how far away the deepest part of the scene is.

Yes even at 2m the errors are already well and alive in the depth data which translates to the flutter and jutter you see in the background of the frame.

The Key issue is how rapidly the D435 scales up the errors - and why the D415 is the only currently viable solution for anything over 1.5m if you want anything close to a stable background.

Westa

[idata]

Did anyone try using D415 in longe range like (7m to 10m)? How good is the point cloud at this distances?

Please share, if you have any images/videos of depth images/point cloud using D415.

Thanks a lot!

[MartyG]

The graph below estimates how depth RMS error is expected to increase with distance on the D415 (green line) and D435 (orange).

The D435 has greater RMS error over distance, though it does also have advantages such as its wider FOV, smaller minimum distance (MinZ) for closer scanning, and faster shutter.

Left-click on the image to see it in full size.