KAIST Urban dataset, position drift when car stops

[FengyuGuo]

Hi. Thanks for your outstanding work. When I test Open VINS in KAIST urban 38 dataset. The trajectory would drift for about several decimeter(~75cm in the picture) even ZUPT is applied. Have any body encounter this phenomena? And any idea to alleviate this?

[goldbattle]

Yes, this is something that I have seen, and have equated it to the fact that the constant velocity of the car and forward motion causes the scale along the forward direction to be very bad (and thus the accelerator biases). These jumps likely are the zero velocity correcting this bad forward velocity / scale / bias, but I am not sure (just a hypothesis). You can also try varying the number of SLAM features to see if that helps.

We have seen that once wheel odometry is included this isn't as bad (zero velocity isn't used in this case). See these works if you are interested in diving deeper:

• http://mars.cs.umn.edu/papers/KejianWu_VINSonWheels.pdf
• https://pgeneva.com/downloads/papers/Lee2020IROS.pdf

[FengyuGuo]

Thanks for you advice. I will check the papers.