A historical perspective on the Kalman filter and its variants
| When |
Mar 16, 2026
from 10:45 to 11:45 |
|---|---|
| Where | Amphi Anne L'Huillier |
| Contact Name | Nelly Pustelnik |
| Attendees |
Audrey Giremus |
| Add event to calendar |
vCal
iCal
|
The Kalman filter is a cornerstone methodology in signal processing and control theory. First deployed for trajectory estimation during NASA’s Appolo program in the early 1960s, it has since found applications across a wide range of scientific and technological domains, including guidance and navigation, flood prediction in hydrology or biological particle tracking, among many others. The algorithm provides recursive online estimates of time-varying parameters from a series of inaccurate measurements. It enjoys strong optimality guarantees under standard assumptions, and is both omputationally efficient and easy to implement, which explains its widespread use. However, the original formulation applies to linear observation and evolution models, disturbed by Gaussian noises. Over the past sixty decades, different approaches have been developed to relax these restrictive assumptions. After revisiting the Kalman filter, this presentation proposes an overview of the most well-known variants: the Extended and Unscented Kalman filters to address non linearities in the models, the Ensemble Kalman filter, particularly well-suited for high dimensional problems, as well as the multiple model approaches for switching dynamics, and the robust Kalman filters to handle non Gaussian noises. Finally, more recent techniques are also introduced, including the coupling with deep learning methods and the extension to non Euclidean variables.