EKF2 local position stops following vision position while platform is moving

[ArkadiuszNiemiec]

I have a vision position system using ZED camera and companion computer connected via ROS. ZED wrapper publishes Odometry message so I wrote a simple script that relays this to mavros PoseStamped topic with rospy.Time.now() in header.
My camera is roughly 20 cm to the front of the Pixhawk.

Steps I did to test it:
Flash 1.8.0 with EKF2, turn off GPS in EKF2_AID_MASK, set EKF2_MAG_TYPE to None calibrate etc. set EKF2_EV_DELAY to 50 ms, EKF2_EVA_NOISE to 0.02 rad and EKF2_EVP_NOISE to 0.01 m.

It looks good when the platform is static, I can see LOCAL_POSITION_NED following the VISION_POSITION_ESTIMATE but when I start moving the platform it diverges, resets and then follows it again.

I have checked frames (ZED publishes in ENU, mavros takes ENU and converts it to NED before publishing mavlink message). How to debug this problem? Is my configuration wrong? ZED sets yaw to 0 when script starts, so maybe there's a problem that vision yaw is not equal to compass yaw? But then I have turned a mag in EKF2..

I have tried to set EKF2_EV_POS_X to +0.2 m but then the vision_position is rejected from the start.

Example logs: https://review.px4.io/plot_app?log=1107b173-7944-4d97-b015-7e8b4cfc2da1

[ArkadiuszNiemiec]

I have forced it to work constantly by increasing EKF2_ev_GATE from 5.0 SD to 10.0 SD.

[mhkabir]

Please post a log with logging enabled from boot.

[tiralonghipol]

Hi. Any news on this? I am having the same behaviour!

[ArkadiuszNiemiec]

@mhkabir I will post it next week, I am out of the office.
@tiralonghipol I made it work by playing with EKF2_EV_GATE but it made the EKF trust vision position even if it is clearly broken. I will try to lower it later and ask @mhkabir to help me debug the problem further.

[tiralonghipol]

yeah, I followed the same steps and the result is the same
I'll let you know if I have any improvement

[ArkadiuszNiemiec]

@mhkabir Ok, I set the EKF2_EV_GATE to the default value and recorded a log from the boot. I was moving the unarmed platform in my arms.

https://logs.px4.io/plot_app?log=4742431c-320c-47b9-8130-c9f93ab75933

[tiralonghipol]

@mhkabir what is your current setup?
are you feeding /mavros/vision_pose/pose with what?

[priseborough]

@ArkadiuszNiemiec

I have the following observations looking at your log:

1. The EKF2_EVP_NOISE parameter setting of 0.01 m is much too low and will lead to the external vision data being rejected at circa 0.05 m of difference to the inertial nav solution. Start at 0.2m.

The ekf assumes the largest of EKF2_EVP_NOISE and eph.epv values and the vehicle_vision_position message in the log has a zero for eph and epv. A recent change pushed to ecl master PX4/PX4-ECL@2354c30 will enable separate horizontal and vertical accuracies when the corresponding PX4/Firmware changes are made.

2. The EKF2_EVA_NOISE value of 0.02 is also too small. It should be no smaller than 0.05 looking at the innovations.

3. If inertial errors continue to dominate, try doubling EKF2_GYR_NOISE and EKF2_ACC_NOISE .

4. The output predictor will track more accurately if the buffer delay is kept short. Currently your buffer size is being determined by two sensors you don't use - GPS and Airspeed. Reduce these to 0 if you are not planning on using them.

5. EKF2_MIN_OBS_DT is currently set to 20 so if the vision position message rate is greater than 50Hz there is a risk of messages being overwritten in the buffer before they are consumed by the ekf.

[ArkadiuszNiemiec]

@priseborough @mhkabir thank you for your help, I have changed few parameters according to your suggestions but I left the EKF2_EV_GATE at 10 SD to make sure that it does not reject vision estmiation.

I have also tuned the UAV to poshold on GPS and I was very happy with results but when flying without a GPS and only on EV it starts to oscilate. Could you tell me how to tune it or is there any other step that I missed? Is my vision data bad or delayed?
Logs: https://logs.px4.io/plot_app?log=97adf845-91a0-4b18-a73c-be4a9d5116d0

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[stale]

Closing as stale.