Online Variational Approximations to non-Exponential Family Change Point Models: With Application to Radar Tracking

Ryan D. Turner, Steven Bottone, Clay J. StanekNeurIPS 2013

In the world of radar tracking, knowing where an object is represents only half the battle. To truly distinguish a Boeing 747 from a flock of birds or random electronic noise, engineers rely on "feature-aided tracking." One of the most critical features used is the Radar Cross Section (RCS), often measured via the Signal-to-Noise Ratio (SNR).

However, the physics of radar dictates that this data rarely follows "nice" Gaussian distributions. Instead, it typically follows a Rice distribution, which is notoriously difficult to handle in real-time Bayesian frameworks.

Radar tracking scenario with crossing targets and clutter

This paper introduces a novel approach to bring the power of Bayesian Online Change Point Detection (BOCPD) to these complex, non-exponential family models using Variational Bayes.

TL;DR

The authors develop a computationally efficient framework for detecting sudden changes in data streams that don't fit standard statistical families (like the Rice distribution in radar). By combining Variational Bayes with Online Change Point Detection, they achieved a 30% improvement in tracking accuracy and significantly reduced "ghost" tracks in crowded airspace.

What Problem Does This Paper Address?

Bayesian Online Change Point Detection (BOCPD) is a gold standard for identifying when the underlying parameters of a process suddenly shift. However, BOCPD has a major catch: it only runs in constant time and memory ($O(1)$) if the underlying model belongs to the exponential family with conjugate priors.

In radar tracking, the RCS feature typically follows a Rice distribution. Because the Rice distribution is not in the exponential family, standard BOCPD becomes computationally explosive for real-time systems. Engineers have historically relied on simpler filters or point estimates (MAP), which struggle to adapt when an aircraft turns or environmental conditions change. This lag leads to "spurious tracks" (tracking objects that don't exist) or losing track of real targets entirely.

Methods

The authors bridge the gap between complex physics and efficient computation through four main innovations:

  1. Variational Bayes (VB) for Rice Distributions: They derived a new VB-EM algorithm specifically for the Rice distribution. By treating the data as being augmented with latent variables, they transformed the "difficult" Rice model into a "conjugate-exponential" form that VB can solve iteratively.
  2. Improved Online VB: Standard online VB often loses accuracy over time or requires re-processing the entire history ($O(n)$). The authors proposed a "feature space" projection method. By approximating the expected sufficient statistics using a finite inner product (a lookup table approach), they kept updates at an $O(1)$ cost while maintaining near-batch accuracy.
  3. The ELBO Proxy: BOCPD requires calculating the "predictive log-likelihood" of new data. Since this is intractable for the Rice distribution, the authors used the Variational Lower Bound (ELBO) as a proxy. This allows the change point mechanism to efficiently "rank" different possible run-lengths.
  4. Censored Data Handling: They integrated a method to handle "missed detections"—situations where the radar scans but the signal is below the detection threshold—ensuring the Bayesian posterior remains accurate even when data is missing.

Key Insights and Ideas

  • The "Aha" Moment on Forgetting: In many online learning tasks, algorithms require "forgetting factors" to ignore old, irrelevant data. The authors realized that when VB is embedded inside a change point model, you don't need these factors. The BOCPD mechanism naturally phases out old data by assigning higher probability to shorter "run-lengths" when a change occurs.
  • RCS as a Discriminator: Using RCS (SNR) isn't just about knowing the target's size; it's about data association. By accurately modeling the Rice distribution, the tracker can tell if a measurement belongs to a known track or is just "clutter" (like birds), even if the physical positions are close together.
  • ELBO as a Valid Proxy: The paper demonstrates that the Variational Lower Bound is a robust enough proxy for the marginal likelihood to drive the change point detection logic effectively.

Key Results

The researchers evaluated their approach using synthetic Rice data, simulated RCS jumps, and real-world flight trajectories from Flightradar24.

RCS RMSE over time showing the proposed method's lower error floor
  • Superior Accuracy: In RCS estimation during simulated jumps, the proposed BOCPD method reached an error floor of 0.56 dB, whereas the industry-standard Interacting Multiple Model (IMM) filter hit a "floor" at 2 dB.
  • Tracking Gains: Using the Single Integrated Air Picture (SIAP) metrics on real flight data, the BOCPD-RCS model showed:
  • 30% improvement in positional accuracy in difficult, high-density scenarios.
  • 11% improvement in velocity accuracy.
  • 6% reduction in the frequency of spurious (false) tracks.
  • Computational Efficiency: Their improved online VB method maintained an error of only $10^{-2}$ nats compared to the batch solution, while naive online updates drifted by as much as 50 nats.

Limitations and Future Work

While the method is powerful, the authors note that the Rice distribution is just one example. The framework is general for non-exponential families, but each new distribution requires its own specific latent variable augmentation strategy to work with VB-EM.

Additionally, the paper focuses on 2D data association, which is computationally manageable. Higher-dimensional assignment problems (tracking many targets across many sensors) remain NP-hard, suggesting that future work could involve casting more complex data association algorithms into this Variational framework to find better approximations.

Summary updated

BibTeX
@inproceedings{turner2013online,
  author = {Ryan D. Turner and Steven Bottone and Clay J. Stanek},
  title = {{Online Variational Approximations to non-Exponential Family Change Point Models: With Application to Radar Tracking}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2013},
}
Figure