[Paper Review] A Generalized Labeled Multi-Bernoulli Filter Implementation using Gibbs Sampling

Hung Gia Hoang, Ba-Tuong Vo+1 more|arXiv (Cornell University)|Jun 2, 2015

Target Tracking and Data Fusion in Sensor Networks28 references3 citations

TL;DR

This paper proposes a joint prediction and update framework for the δ-Generalized Labeled Multi-Bernoulli (δ-GLMB) filter using Gibbs sampling, eliminating redundant truncations in conventional implementations. By integrating prediction and update into a single step and leveraging stochastic sampling to identify high-weight components, the method reduces computational complexity from cubic to linear in measurement count while maintaining or improving tracking accuracy, especially under high-clutter conditions.

ABSTRACT

This paper proposes an efficient implementation of the generalized labeled multi-Bernoulli (GLMB) filter by combining the prediction and update into a single step. In contrast to the original approach which involves separate truncations in the prediction and update steps, the proposed implementation requires only one single truncation for each iteration, which can be performed using a standard ranked optimal assignment algorithm. Furthermore, we propose a new truncation technique based on Markov Chain Monte Carlo methods such as Gibbs sampling, which drastically reduces the complexity of the filter. The superior performance of the proposed approach is demonstrated through extensive numerical studies.

Motivation & Objective

• To address the computational inefficiency of the traditional two-stage δ-GLMB filter implementation, which performs separate truncations in prediction and update steps.
• To reduce redundant computation caused by processing low-weight components through intermediate truncations in the original approach.
• To develop a unified framework that combines prediction and update into a single step, enabling more efficient truncation.
• To introduce a stochastic truncation method using Gibbs sampling that adaptsively selects significant components without sorting.
• To demonstrate superior performance and reduced runtime compared to deterministic ranked assignment methods, especially in high-clutter environments.

Proposed method

• Proposes a joint prediction and update formulation that eliminates intermediate truncation by directly linking filtering densities across time steps.
• Uses a single ranked assignment algorithm per iteration to truncate the filtering density after joint prediction and update, replacing two separate truncations.
• Introduces Gibbs sampling as a stochastic alternative to deterministic ranked assignment for identifying high-weight δ-GLMB components.
• Employs Markov Chain Monte Carlo (MCMC) sampling to generate significant components based on their statistical weights, avoiding sorting overhead.
• Leverages the direct relationship between predicted and updated components to reduce computational redundancy and improve efficiency.
• Applies the method to both full δ-GLMB and approximated Labeled Multi-Bernoulli (LMB) filters, showing broad applicability.

Experimental results

Research questions

• RQ1 Can a joint prediction and update framework reduce computational complexity in δ-GLMB filtering compared to the traditional two-stage approach?
• RQ2 Does Gibbs sampling-based truncation outperform deterministic ranked assignment in terms of accuracy and runtime under high-clutter conditions?
• RQ3 Can the proposed method maintain or improve tracking performance while reducing the number of required computations?
• RQ4 How does the stochastic nature of Gibbs sampling affect component selection compared to deterministic sorting in high-dimensional multi-object tracking?
• RQ5 Is the joint implementation scalable and effective for real-time applications with high false alarm rates?

Key findings

• The joint prediction and update approach reduces computational complexity by eliminating redundant truncations, achieving 1 to 2 orders of magnitude faster execution than the original two-stage method.
• The Gibbs sampling-based truncation method reduces complexity from cubic to linear in the number of measurements, significantly improving scalability.
• Under high-clutter conditions (100 false alarms per scan), the Gibbs sampler achieves lower OSPA distance from time step k=75 onward, indicating better tracking accuracy.
• The joint approach maintains equivalent cardinality estimation performance to the traditional method, with mean estimates and standard deviations closely matching over 100 Monte Carlo trials.
• The Gibbs sampler outperforms the ranked assignment method in identifying new-born targets in difficult scenarios due to its randomized exploration of high-weight components.
• The proposed method is applicable to approximations such as the Labeled Multi-Bernoulli (LMB) filter, extending its utility beyond the full δ-GLMB model.