State smoothing in Markov jump systems with lagged mode observation

Yan Liang, Lei Zhang, Quan Pan, Tongwen ChenInternational Journal of Adaptive Control and Signal Processing 2010 · DOI

TL;DR

This paper tackles the challenge of tracking targets using Markov Jump Systems (MJSs) where crucial "mode" information (like target orientation) is delayed compared to the primary state measurements. They propose a novel recursive smoothing algorithm that effectively incorporates this lagged mode data to improve tracking accuracy, outperforming existing methods in simulations.

What Problem Does This Paper Address?

Many real-world tracking scenarios involve systems that can switch between different operational modes (e.g., a target changing its movement pattern). Markov Jump Systems (MJSs) are a powerful framework for modeling these systems, where both the system's continuous state (like position and velocity) and its discrete mode evolve over time. Traditionally, estimators for MJSs assume that measurements of the system's state and its current mode arrive at the same time. However, in applications like image-based target tracking, obtaining the mode information (e.g., identifying the target's orientation from an image) can be computationally intensive and thus delayed relative to the primary state measurements (e.g., radar range and bearing). This delay can degrade tracking performance. This paper addresses this gap by developing a method to effectively utilize state measurements and lagged mode observations for improved state estimation in MJSs.

Methods

The core of this paper is a novel recursive estimator designed for MJSs with lagged mode observations. The authors employ a state augmentation technique, where the current state and the previous state are combined into a larger state vector. This allows them to transform the problem into a standard Kalman filtering framework for each mode. The key innovation lies in how they recursively derive the conditional state mean and conditional model probabilities by incorporating both the delayed mode observations and the current state measurements. The proposed algorithm involves several steps, including:

  • First-stage model probability update: Estimating the probability of each mode based on past information.
  • Interaction of estimates of sub-filters: Combining information from individual mode-specific filters.
  • Estimation of sub-filter: Performing Kalman filtering for each mode.
  • Second-stage model probability update: Refining mode probabilities using the latest measurements.
  • Fusion of estimates of sub-filters: Combining the estimates from all modes to produce the final smoothed state estimate.

The authors also introduce an assumption that each reinitialized state estimate can be approximated by a Gaussian distribution, which simplifies the calculations and makes the estimator computationally feasible.

Datasets

Not discussed in the paper. The authors use simulated data for their experiments.

Key Insights and Ideas

  • Lagged mode information is valuable: The paper strongly demonstrates that even delayed information about the system's mode can significantly improve tracking accuracy, especially when the mode observations are of high quality.
  • State augmentation simplifies the problem: By augmenting the state vector, the authors can leverage well-established Kalman filtering techniques for each mode, making the derivation more manageable.
  • Recursive approach for efficiency: The proposed estimator is recursive, meaning it updates its estimates sequentially as new data arrives, which is crucial for real-time applications.
  • OFL-IMM as a special case: The paper highlights that their proposed method reduces to the standard One-step Fixed-Lagged IMM (OFL-IMM) algorithm when the mode observations provide no useful information, demonstrating the generality of their approach.

Key Results

The proposed smoother was evaluated on maneuvering target tracking scenarios. The simulations showed that:

  • The proposed smoother significantly improves tracking performance (both position and velocity estimation) compared to the IE-IMM (Image-Enhanced IMM) algorithm and the OFL-IMM algorithm.
  • The improvement is more pronounced when the mode observations are of higher quality.
  • The computational cost of the proposed method is slightly higher than OFL-IMM but offers a substantial gain in accuracy. Compared to IE-IMM (a filtering approach), the smoothing method provides better accuracy at the cost of increased computation.

Limitations and Future Work

  • Gaussian Approximation: The paper relies on the assumption that state estimates can be approximated by Gaussian distributions. While this is a common simplification in MJS estimation, it might limit accuracy in highly non-Gaussian scenarios. The authors suggest that introducing particle filters could remove this assumption but at a significant computational cost.
  • Computational Cost: While the proposed method offers improved accuracy, it is more computationally intensive than filtering methods like IE-IMM. The authors acknowledge that state smoothing is inherently more complex than state filtering.
  • Future Work: The authors suggest that the assumption of Gaussian approximation can be removed by using particle filters, which would allow for more accurate estimation in non-Gaussian scenarios, albeit with higher computational demands.

Summary updated

Abstract

Abstract Estimation involving Markov jump systems (MJSs) is widely used in target tracking, speech recognition and communication. It is assumed in MJSs that state measurement and mode observation are synchronous. In applications such as image‐based target tracking, the target orientation, as one of the mode observations, needs additional computation time for pattern recognition and thus can be delayed. This motivates us to explore the smoothing problem of MJSs with mode observation lagged to state measurement. This brief paper presents a recursive estimator by deriving the conditional state mean and the conditional model probability from both delayed mode observation and state measurement. Simulations on maneuvering target tracking are carried out to validate the performance of the proposed smoother in comparison with existing methods. Copyright © 2010 John Wiley & Sons, Ltd.
BibTeX
@article{liang2010state,
  author = {Yan Liang and Lei Zhang and Quan Pan and Tongwen Chen},
  title = {{State smoothing in Markov jump systems with lagged mode observation}},
  journal = {International Journal of Adaptive Control and Signal Processing},
  year = {2010},
  doi = {10.1002/acs.1168},
}
Figure