Acoustic Source Tracking Based on Probabilistic Data Association and Distributed Cubature Kalman Filtering in Acoustic Sensor Networks

Yang Chen, Yideng Cao, Rui Wang2022 · DOI

TL;DR

This paper introduces a novel distributed acoustic sensor network approach for tracking sound sources, even in challenging noisy and reverberant conditions. By combining probabilistic data association with a distributed cubature Kalman filter and a smart fusion strategy, the method achieves robust and accurate tracking, outperforming existing distributed methods and rivaling centralized approaches.

What Problem Does This Paper Address?

Accurately tracking the location of a sound source, like a speaker, is crucial for many applications, from teleconferencing to human-computer interaction. Traditional methods often rely on microphone arrays with specific geometric arrangements and centralized processing, which can be inflexible and prone to single points of failure. Distributed acoustic sensor networks offer a more adaptable solution, but they face significant challenges in real-world environments. Noise and reverberation can distort the sound signals, leading to unreliable measurements and making it difficult to distinguish the true source from echoes or background sounds. Existing distributed tracking methods often struggle with these adverse conditions, as they may rely on single, potentially erroneous, peak measurements from sound correlation. This paper aims to address these limitations by developing a more robust and accurate distributed tracking system that can effectively handle noise and reverberation.

Methods

The core of this work lies in the proposed Probabilistic Data Association-based Distributed Cubature Kalman Filter (PDA-DCKF). Here's a breakdown of the key components:

  • Probabilistic Data Association (PDA): Instead of relying on a single "best guess" observation from each microphone node, PDA allows the system to consider multiple potential observations. This is crucial in noisy environments where the strongest signal might not be the true source. PDA helps to probabilistically associate these multiple observations with the actual sound source, improving robustness.
  • Cubature Kalman Filter (CKF): CKF is an efficient and accurate method for nonlinear filtering, which is well-suited for acoustic tracking where the relationship between source location and microphone signals is nonlinear. The paper first develops a single-node PDA-CKF.
  • Distributed Architecture: The PDA-CKF is then extended to a distributed network (PDA-DCKF). In this setup, each node in the network processes its local observations using PDA-CKF. Crucially, nodes also exchange information with their neighbors, allowing for a more comprehensive understanding of the acoustic environment.
  • Fusion Strategy: A novel fusion strategy is introduced to combine the local estimates from each node into a global estimate. This strategy intelligently weights the contributions of each node based on two factors:
  • Received Signal Energy: Nodes receiving stronger signals are given more weight.
  • Mean Square Error (MSE) of Position Estimation: Nodes with more accurate local estimates (lower MSE) are also prioritized. This dynamic weighting ensures that more reliable local information contributes more to the final global tracking result.
  • Langevin Model: The paper uses the Langevin model to describe the dynamic motion of the acoustic source, which is a standard approach for modeling target movement in such scenarios.
  • TDOA Estimation: Time Difference of Arrival (TDOA) is used as the primary observation for tracking. The paper details a method to extract multiple TDOA candidates from the generalized cross-correlation function (GCC) to overcome issues with spurious peaks.

Datasets

The paper does not use pre-existing datasets. Instead, it relies on simulated data generated within a controlled laboratory environment. The simulation setup includes:

  • A typical room environment (6m x 6m x 3m).
  • A distributed acoustic sensor network of 12 nodes, each with two microphones.
  • Simulated speech signals with varying Signal-to-Noise Ratios (SNR) and reverberation times (T60).
  • Two distinct acoustic source trajectories: a linear path and a nonlinear (semicircle) path.
  • Room Impulse Responses (RIRs) simulated using the image-source method to mimic realistic acoustic environments.

Key Insights and Ideas

  • The Power of Multiple Observations: The paper strongly demonstrates that moving beyond single-peak TDOA measurements and embracing multiple potential observations via PDA is critical for robust tracking in noisy and reverberant conditions.
  • Distributed Intelligence with Smart Fusion: Simply distributing processing isn't enough. The proposed dynamic weighting fusion strategy, considering both signal energy and local estimation accuracy, is key to achieving near-centralized performance from a distributed network.
  • Robustness to Network Damage: The PDA-DCKF shows remarkable resilience even when some nodes in the network are "damaged" (i.e., unable to communicate), highlighting the inherent fault tolerance of the distributed approach.
  • Bridging the Gap to Centralized Performance: The PDA-DCKF achieves tracking performance very close to that of a centralized approach (CCKF), but with significantly lower computational cost and network burden, making it more practical for real-world deployment.

Key Results

  • Superior Performance in Noise and Reverberation: The PDA-DCKF consistently outperforms other distributed methods (DCKF, DUKF, DEKF, DICKF) in terms of Root Mean Square Error (RMSE) across various SNR and reverberation conditions.
  • Near-Centralized Accuracy: The proposed method achieves tracking accuracy comparable to the centralized CKF (CCKF), especially at higher SNRs.
  • Robustness to Node Failure: Even with node damage (e.g., one or two nodes failing), the tracking accuracy degrades only slightly, demonstrating good robustness.
  • Effective Nonlinear Trajectory Tracking: The algorithm successfully tracks a nonlinear (semicircle) trajectory, indicating its capability beyond simple linear movements.

Limitations and Future Work

  • Computational Cost: While more efficient than centralized methods, distributed filtering can still be computationally intensive. The paper mentions that particle filter-based methods, which are also used for acoustic tracking, have even higher computational costs, limiting their real-time application.
  • Multi-Source Tracking: This paper focuses on tracking a single acoustic source. The authors acknowledge that multi-source tracking is a more complex problem and is beyond the scope of this work.
  • Future Work: The paper implicitly suggests that further optimization of the fusion strategy and exploration of more advanced data association techniques could lead to even better performance. Extending the method to multi-source scenarios would be a natural next step.

Summary updated

BibTeX
@misc{chen2022acoustic,
  author = {Yang Chen and Yideng Cao and Rui Wang},
  title = {{Acoustic Source Tracking Based on Probabilistic Data Association and Distributed Cubature Kalman Filtering in Acoustic Sensor Networks}},
  year = {2022},
  doi = {10.3390/s22197160},
}
Figure