Any suggestions how to improve SLAM?

[HWade]

Hi, in my experiments with the RS Slam module I noticed, that the drift is sometimes quite strong.

I assume that Intel has implemented some kind of Kalman Filter (didnt check the source code yet) which means, there should be some initialialization step for the gravity vector estimation and probably the scaling of the IMU.

Would it help to improve the accuracy, if I rotate the camera and hold still while some any of the three axis is showing towards the gravity vector?

Thanks in advance!

[MartyG]

The experts on RealSense SLAM will likely be around here on the forum from Monday onward, unless they are reading today in their free time. In the meantime, you might find some interest in a recent discussion about resetting drift if you haven't seen it already. gGod luck!

[idata]

Hi Hanneswader,

Thanks for contacting us. Did you check the thread provided by MartyG? Additionally, please take a look at the "Known limitations" section on this site: https://software.intel.com/sites/products/realsense/release_notes/slam_release_notes.html Release Notes for Intel® RealSense™ SLAM Library in order to take that into consideration for your tests. Regards,-Yermi A.

[HWade]

Hi,

thank you very much for the hint on the relocalization. Currently, I am implementing relocalization on my own using some experimental ideas... However, I rely on the VINS odometry provided by the RealSense. Thus, the better the odometry, the better the final result.

The "known limitation" items gave me at least some hint on the issue.

Thanks again