Particle Flow Gaussian Particle Filter

Karthik Comandur, Yunpeng Li, Santosh Nannuru

TL;DR

The Particle Flow Gaussian Particle Filter (PFGPF) is a new state estimation algorithm that combines the robustness of Gaussian Particle Filters with the precision of invertible particle flow. By migrating particles toward high-probability regions using physical flow equations, it overcomes the "weight degeneracy" problem in high-dimensional spaces without needing a separate resampling step.

What Problem Does This Paper Address?

In fields like robotics, radar tracking, and finance, we often need to estimate the "state" of a system (like the position of a drone) based on noisy, non-linear measurements. Particle Filters (PF) are the gold standard for this, using a "cloud" of particles to represent possible states.

However, standard PFs face a massive hurdle called weight degeneracy: in high-dimensional spaces or when measurements are very precise, almost all particles end up with a weight of zero. This makes the filter fail. While "resampling" can help, it often leads to "sample impoverishment," where the filter loses diversity and becomes overconfident in a wrong guess.

The authors aim to create a filter that works in high dimensions (where the "curse of dimensionality" usually kills particle filters) by intelligently moving particles toward the most likely state before weighting them.

Methods

The authors introduce the PFGPF, which merges two existing concepts:

  1. Gaussian Particle Filter (GPF): A variant that approximates the system's state distribution as a Gaussian (Bell curve). This avoids the need for resampling because you can simply re-draw particles from the updated Gaussian at each step.
  2. Invertible Particle Flow: Instead of just guessing where the state might be, this method uses a differential equation (derived from the Fokker-Planck equation) to "flow" particles from the prior distribution toward the posterior (the state after seeing new data).

The Secret Sauce: By using invertible flow (specifically the Localized Exact Daum and Huang, or LEDH, method), the authors can mathematically track how the particle density changes during the flow. This allows them to calculate a "modified importance sampling weight," ensuring the filter remains statistically consistent while placing particles exactly where they are most useful.

Datasets

The researchers evaluated their algorithm using two complex numerical simulations:

  • Multi-target Acoustic Tracking: Tracking 4 independent targets in a 16-dimensional state space using 25 acoustic sensors. This setup is challenging because the sensor measurements are highly non-linear.
  • Large Spatial Sensor Networks: A massive 144-dimensional state space involving a "Skewed-t" dynamic model (which has "heavy tails" to simulate extreme behavior) and Poisson count measurements. This test specifically targets the algorithm's performance in very high dimensions.

Key Insights and Ideas

  • Natural Synergy: Particle flow and Gaussian filters are a perfect match because both rely on Gaussian assumptions to simplify complex math. Integrating them allows for a more "principled" way to generate proposal distributions.
  • No Resampling Needed: Because the GPF framework approximates the posterior as a Gaussian, the "sample impoverishment" seen in traditional particle filters is bypassed entirely.
  • Mathematical Rigor: Unlike some "particle flow filters" that lack statistical consistency, the PFGPF uses the encompassing importance sampling mechanism. This means it inherits the asymptotic properties of the GPF—essentially, as you add more particles, the filter is guaranteed to converge to the true value.

Key Results

  • Superior Accuracy: In the acoustic tracking simulation, PFGPF achieved the lowest error (0.73m OMAT error with 500 particles), significantly outperforming standard particle flow filters (EDH/LEDH) and other hybrid filters.
  • Stability in High Dimensions: In the 144-dimensional sensor network test, the PFGPF showed its strength over time. While other filters' performance degraded over 30 time steps, the PFGPF maintained a much lower Mean Squared Error (0.87 vs. 1.08 for the next best hybrid).
  • Efficiency: While more computationally intensive than a simple flow filter, the PFGPF's execution time was comparable to other state-of-the-art hybrid filters, making it a viable candidate for real-world high-dimensional tracking.

Limitations and Future Work

The main limitation is the Gaussian assumption. If the true state of the system is "multi-modal" (meaning there are two or more equally likely but different possibilities, like a target that could have turned either left or right), a single Gaussian curve cannot represent it accurately.

Future Work: The authors suggest extending this framework to Gaussian Sum Particle Filters. This would allow the algorithm to use multiple "Bell curves" simultaneously, potentially handling complex, multi-modal scenarios while retaining the high-dimensional benefits of particle flow.

Summary updated

Abstract

State estimation in non-linear models is performed by tracking the posterior distribution recursively. A plethora of algorithms have been proposed for this task. Among them, the Gaussian particle filter uses a weighted set of particles to construct a Gaussian approximation to the posterior. In this paper, we propose to use invertible particle flow methods, derived under the Gaussian boundary conditions for a flow equation, to generate a proposal distribution close to the posterior. The resultant particle flow Gaussian particle filter (PFGPF) algorithm retains the asymptotic properties of Gaussian particle filters, with the potential for improved state estimation performance in high-dimensional spaces. We compare the performance of PFGPF with the particle flow filters and particle flow particle filters in two challenging numerical simulation examples.
BibTeX
@inproceedings{comandur2022particle,
  author = {Karthik Comandur and Yunpeng Li and Santosh Nannuru},
  title = {{Particle Flow Gaussian Particle Filter}},
  booktitle = {2022},
  year = {2022},
  doi = {10.23919/fusion49751.2022.9841354},
}

Loading related content...

Figure