Efficient parametric non-Gaussian dynamical filtering

Filtering is a key component of many modern control Systems: from noisy measurements, we want to be able to determine the state of some system as it evolves over time. Modern applications that require filtering tend to implement a filter from one of two main families of techniques: the Kalman filter (and associated extensions) and the particle filter. Each is popular and correct in its own right for certain applications, however each also has its limitations making it unsuitable for other applications. In this paper we propose a new filter based on the Student-t distribution to address the problems of the aforementioned filters: a filter which admits multimodal state hypotheses, is more robust to outliers, and remains computationally tractable in high-dimensional spaces.