ZED2 Camera Tracking pose_smoothing


I’m currently trying to use the ZED2 as a camera tracker. Which works really well!
The only little thing that bothers me is that she doesn’t correct the pose drift softly, but rather jumps.

Even if I set

    tracking_parameters = sl.PositionalTrackingParameters()
    tracking_parameters.enable_pose_smoothing = True
    tracking_parameters.set_floor_as_origin = False
    tracking_parameters.enable_area_memory = True

    if self.load_old_map == True:
        tracking_parameters.area_file_path = "****"

    if err != sl.ERROR_CODE.SUCCESS:

    zed_pose = sl.Pose()
    runtime_parameters = sl.RuntimeParameters()

, it jumps like in the video. https://youtu.be/4VBnCsurJxk

I know that backlighting is not optimal. However, the position also jumps in other environments.

I also tried to set every value to FASLE, whereby the tracker reacts exactly the same.
So maybe there is a mistake in my Code?

Is there a solution? Or is it simply not possible to prevent that?

Thank you!

Hi @ScaredyStefan
can you verify that the IMU of your camera is correctly calibrated?
You can do that by opening “ZED Sensor Viewer” and checking that there are no drifts when selecting the camera “Orientation” plot.


I can’t see any drift.
What I find suspicious, is how the spatial mapping behaves. https://youtu.be/N-BAcJ9Hi_4 (0:44)
It does not recognize straight walls and other stuff, but i have very little experience in this topic.
I have uploaded a recording from die IMU and the json from the diagnostic tool … I hope it helps


20210429_160437-ZED2-sn_21974610-ZEDSensorsData.csv (884.7 KB) ZED_Diagnostic_Results.json (19.6 KB)


today I did some “research” again. What I noticed is that the tracking confidence often drops to zero for a few frames.

Now 2 questions:

  1. What would an optimal environment for a good tracking would look like? Maybe I can test in such an environment.

  2. What are the best settings for the tracking? So when I definitely only need the pose data.

  1. The ZED devices are passive stereo cameras, so they need good visual features to provide a good depth estimation and so a good positional tracking elaboration. Walls and surfaces with a uniform color are correctly processed but do not always provide good information that allows getting precise results. An environment rich in texture is the best place for a ZED. The internal IMU allows a better estimation of the camera attitude, but when the visual information lacks the noises increase.

  2. See 1