Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning IMM Filter Parameters from Measurements using Gradient Descent

André Brandenburger, Folker Hoffmann, Alexander Charlish

TL;DR

The paper tackles automatic tuning of Interacting Multiple Model (IMM) filter parameters using only measurement data, removing the need for ground-truth. It introduces a differentiable framework that minimizes a measurement-likelihood loss ${\mathcal{L}}({\bm{\theta}})$ via gradient descent to update the IMM's mode transitions, process noises, and measurement parameters. Through simulated experiments with two motion modes, the learned IMM matches the performance of an IMM with true parameters and outperforms a single-mode Kalman filter, while providing interpretable parameter updates. This approach enables practical, measurement-driven adaptation of complex multi-mode trackers for real sensors and can extend to more sophisticated architectures.

Abstract

The performance of data fusion and tracking algorithms often depends on parameters that not only describe the sensor system, but can also be task-specific. While for the sensor system tuning these variables is time-consuming and mostly requires expert knowledge, intrinsic parameters of targets under track can even be completely unobservable until the system is deployed. With state-of-the-art sensor systems growing more and more complex, the number of parameters naturally increases, necessitating the automatic optimization of the model variables. In this paper, the parameters of an interacting multiple model (IMM) filter are optimized solely using measurements, thus without necessity for any ground-truth data. The resulting method is evaluated through an ablation study on simulated data, where the trained model manages to match the performance of a filter parametrized with ground-truth values.

Learning IMM Filter Parameters from Measurements using Gradient Descent

TL;DR

The paper tackles automatic tuning of Interacting Multiple Model (IMM) filter parameters using only measurement data, removing the need for ground-truth. It introduces a differentiable framework that minimizes a measurement-likelihood loss via gradient descent to update the IMM's mode transitions, process noises, and measurement parameters. Through simulated experiments with two motion modes, the learned IMM matches the performance of an IMM with true parameters and outperforms a single-mode Kalman filter, while providing interpretable parameter updates. This approach enables practical, measurement-driven adaptation of complex multi-mode trackers for real sensors and can extend to more sophisticated architectures.

Abstract

The performance of data fusion and tracking algorithms often depends on parameters that not only describe the sensor system, but can also be task-specific. While for the sensor system tuning these variables is time-consuming and mostly requires expert knowledge, intrinsic parameters of targets under track can even be completely unobservable until the system is deployed. With state-of-the-art sensor systems growing more and more complex, the number of parameters naturally increases, necessitating the automatic optimization of the model variables. In this paper, the parameters of an interacting multiple model (IMM) filter are optimized solely using measurements, thus without necessity for any ground-truth data. The resulting method is evaluated through an ablation study on simulated data, where the trained model manages to match the performance of a filter parametrized with ground-truth values.
Paper Structure (9 sections, 13 equations, 3 figures, 2 tables)

This paper contains 9 sections, 13 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Interacting Multiple Model Filter
  3. IMM Parameter Optimization
  4. Experimental Results
  5. Data Generation
  6. Loss Evaluation
  7. Training
  8. Comparison of IMM and Single-Mode Kalman Filter
  9. Conclusion

Figures (3)

  • Figure 1: An overview of the IMM filter optimization strategy. Based on its parametrization ${\bm{\mathrm{\theta}}}$, the IMM filter generates tracks (top) which are utilized by the gradient descent optimizer (bottom). The optimizer successively updates the IMM filter parameters with the goal to minimize the negative log-likelihood of the measurements.
  • Figure 2: A sample from the datasets. The ground-truth trajectory is shown on the left and starts at the coordinate origin. The mode switch is highlighted by the red circle. The ground-truth target motion mode and the absolute acceleration (process noise) are displayed on the right.
  • Figure 3: A projection of the loss for different parameters. For each graph, a single variable is assessed for a fixed dataset. The corresponding ground-truth value of the respective IMM parameter in the analyzed dataset is marked red. The loss with respect to the IMM parameter is shown in blue and the posterior RMSE to the ground-truth trajectories is shown in orange.