I have just tested the location tracking provided by the Zed 2i, but the effect is not very satisfactory:
First, if the device is completely stationary, the position can be completely unchanged, which is very good.
However, once the device starts shaking (even if the shaking is very small, such as jumping a few times on the floor next to the device), position tracking exposes a problem: As long as there are larger objects passing through the view, position tracking can generate very large errors. For example, when we ask a person to walk 30 centimeters in front of the device, the device can only see the background, ground, and the person’s legs. After the person walks past, position tracking generates an error of 50 centimeters, and when the person leaves, the device cannot return to its original position based on spatial memory (set “enable_area_memory” to true in the tracking parameter).
“I thought that the IMU built into the device could help position tracking maintain stability in this situation, but the current effect seems to be poor. What can I do to reduce the drift caused by this dynamic interference in close range?”?
My scene is very simple, except for some robots (occasionally very close), which are all static objects. 
Hi @yuanlu
we are releasing in the next few days a new version of the ZED SDK with an improved Positional Tracking module.
Hello, do you have a specific time? We officially use it on April 14th, so we don’t know if we can make it in time.
Released yesterday.
If you need the ROS or ROS 2 wrappers you must wait a few hours yet
I tried it on other devices and the results were good.
Unfortunately, my Jetson NX has been upgraded to CUDA12.1, and I look forward to the new version being released
CUDA 12.1 on the Jetson NX? Are you sure of what you say?
The NVIDIA Jetpack provides CUDA 11.4 in the latest version…
Hmmm, I have already encountered a problem here. I thought it was supported. Previously, upgrading recklessly caused nx to be unusable, and it has been flushed back to 11.4…