Abstract:

We consider the joint state and measurement- noise parameter estimation problem for nonlinear state-space models with asynchronous, variable-rate, and independent measurement sources. We approach the problem using variational Bayes Kalman filters (VB-KFs). By leveraging that the measurements from different sources are independent, we develop an asynchronous VB-KF (AVB-KF), which processes measurements from different sources sequentially and at a variable rate. Hence, in the measurement update step, we only update the noise parameters of measurements that have been processed at a particular time step. This results in faster computations, especially as the measurement dimension and the number of sensors grow. We validate the approach on a realistic application of autonomous mobile-robot platooning, where we perform fusion of multiple sensor modalities with time-varying noise characteristics. The results indicate more than a factor of two improvements measured as a time-averaged absolute error compared to a nonadaptive implementation.