[Paper Review] The Adaptive Labeled Multi-Bernoulli Filter

Andreas Danzer, Stephan Reuter+1 more|arXiv (Cornell University)|Dec 20, 2018

Target Tracking and Data Fusion in Sensor Networks19 references5 citations

TL;DR

This paper proposes the Adaptive Labeled Multi-Bernoulli (ALMB) filter, which dynamically switches between the computationally efficient Labeled Multi-Bernoulli (LMB) and the more accurate δ-Generalized Labeled Multi-Bernoulli (δ-GLMB) representations based on Kullback-Leibler divergence and entropy. The ALMB filter achieves near-δ-GLMB accuracy in complex scenarios while maintaining LMB-level efficiency in simpler cases, outperforming both in tracking precision and computational load.

ABSTRACT

This paper proposes a new multi-Bernoulli filter called the Adaptive Labeled Multi-Bernoulli filter. It combines the relative strengths of the known Delta-Generalized Labeled Multi-Bernoulli and the Labeled Multi-Bernoulli filter. The proposed filter provides a more precise target tracking in critical situations, where the Labeled Multi-Bernoulli filter looses information through the approximation error in the update step. In noncritical situations it inherits the advantage of the Labeled Multi-Bernoulli filter to reduce the computational complexity by using the LMB approximation.

Motivation & Objective

• Address the trade-off between computational complexity and tracking accuracy in multi-object tracking systems.
• Overcome the LMB filter's loss of information in ambiguous data association scenarios due to approximation errors in the update step.
• Leverage the δ-GLMB filter’s superior accuracy in critical situations while preserving the LMB filter’s low computational cost in non-critical scenarios.
• Develop an adaptive mechanism that automatically selects the optimal representation (LMB or δ-GLMB) based on statistical criteria without manual tuning.

Proposed method

• Use Kullback-Leibler (KL) divergence and entropy as switching criteria to detect when the LMB approximation introduces significant error.
• Represent the multi-object posterior as a mixture of LMB and δ-GLMB components, maintaining a subset of tracks in δ-GLMB form only when needed.
• Apply the KL divergence between the LMB approximation and the true posterior to assess approximation quality and trigger adaptive switching.
• Use entropy as a secondary criterion to detect multi-modal cardinality distributions, indicating high uncertainty in object count.
• Integrate the adaptive switching into a modular filter architecture that supports parallel processing of independent track groups.
• Implement the filter using Gaussian mixture (GM) approximations for spatial distributions and apply standard prediction, update, pruning, and track extraction steps.

Experimental results

Research questions

• RQ1 Can an adaptive switching mechanism between LMB and δ-GLMB representations improve tracking accuracy without sacrificing computational efficiency?
• RQ2 How can statistical measures like KL divergence and entropy be used to detect critical tracking scenarios requiring higher-fidelity representation?
• RQ3 Does the ALMB filter maintain the low computational cost of the LMB filter in non-critical scenarios while achieving δ-GLMB-level accuracy in challenging situations?
• RQ4 Can the proposed switching criteria be applied independently of the spatial distribution representation, enabling use in diverse implementations like SMC or Gamma-Gaussian-Inverse-Wishart?

Key findings

• In non-critical scenarios (e.g., time k < 55), the ALMB filter performs identically to the LMB filter, confirming its ability to maintain low computational load.
• In critical scenarios with ambiguous data association (e.g., time k > 50 ∧ k < 65), the ALMB filter outperforms the LMB filter by preserving track identity and reducing label switching.
• The ALMB filter achieves OSPAT distances comparable to the δ-GLMB filter in challenging situations, demonstrating near-optimal accuracy.
• Computation time of the ALMB filter remains nearly identical to the LMB filter, with only a slight increase in critical regions, while still outperforming the δ-GLMB filter in those intervals.
• In a high-density scenario with up to 16 objects, the ALMB filter consistently matches or exceeds LMB performance, confirming robustness under complex dynamics.
• The modular design and use of general-purpose criteria (KL divergence and entropy) make the ALMB framework extensible to other implementations, including SMC and Gamma-Gaussian-Inverse-Wishart.