We use the Snapdragon Flight together with ROS (https://github.com/ATLFlight/snap_vio) for obtaining a pose. However, when using it on our drone, it often happens that the values of the mvVISLAM pose are largely increasing or decreasing (for example after 2 or 3 seconds we obtain a postion (x=200, y=400, z=500)).
Usually, this goes hand in hand with "Intrinsic camera cal questionable" errors, but often there is not even any error code returned. We already tried default fisheye camera calibration and our own calibration. We also noticed that the IMU values during flight are very noisy, but havent found a way to set a parameter for that. What could we do to obtain a reliable pose?
You did not mention the type of robot you are using but it sounds like you are not getting VISLAM properly initialized before proceeding, you have an overly noisy IMU, or very bad camera calibration. If you did your own camera calibration and the undistorted image looked at all reasonable, that is unlikely to be the problem. If you have a very different placement of the camera relative to the IMU than the example code initialization shows then you need to measure the distances and rotations again and change the default values to mvVISLAM_Initialize(); measurements from a CAD drawing should be close enough.
The example code that comes with MV or and the snap_vio code both assume a UAV sitting on landing gear with the camera pointing out at the ground for initialization. That is one way to get a good initialization. Regardless of how good the initialization is, the other way to get a nicely converged VISLAM system is to have very observable acceleration/gyro values. This can be done with far more aggressive takeoffs or hand motions beforehand.
In terms of IMU noise, you can increase the accel noise settings but that slows down filter convergence. Therefore, that should more just be tried during diagnosing the problem and is poor long-term solution. The best way to improve IMU noise is to add mechanical dampening between the board with the IMU + camera on it and the rest of the UAV which has the vibrating motors on it.