Very noisy results with IMU factors.

[ג'וש אלגרבלי‎]

Hey all.

I have a visual odometry application that uses projection factors to optimise the body pose.

The body frame is FLU, the camera frame is RDF.

I also have IMU data recived in FLU->ENU frame at about 50hz.

I worked untill now only with GenericProjectionFactors which gave more than decent results.

Because I have an Imu in my system I wanted to add the Imu into the graph. 

My setup as of now(I know it is not optimal) is frame by frame.

I added the Imu factors as I saw in some examples. 

I set the accel covariance and gyro covariance according to the manufacturer’s data sheet.

And set the integration covariance to about 1e-6.

I assume constant bias, which I assume is pretty ok because of my frame by frame setup.

What I see is happening is that the output becomes worse and very noisy compared to GT and compared to when I only use the Generic projection factors.

I did see a little improvement regarding the understanding of the direction of motion, especially in pure vertical motion, but still very noisy.

No matter how much I played with:

Accel covariance 

Gyro covariance 

Integration covariance

The output dosent change one bit.

What I see that affects the most is:

The velocity noise and the bias noise.

I couldn’t manage to find the sweet spot.

I don’t understand what am I doing wrong.

Maybe my coordinate frames are wrong?

Thanks in advanced for any help.