Revision history [back]

The Mahalanobis thresholds are going to use the covariance of your state and measurement for the variables in question. We don't publish the linear acceleration covariance, but I would assume that they are quite small, and so I think that even a moderate amount of noise in the sensor is going to require large rejection thresholds.

Really, those rejection parameters can probably be removed (i.e., left to their defaults of numeric_limits<double>::max()) until you have a firm understanding of your sensor error and error growth in your state estimate.