Multiple scan data association by convex variational inference

Data association, the reasoning over correspondence between targets and measurements, is a problem of fundamental importance in target tracking. Recently, belief propagation (BP) has emerged as a promising method for estimating the marginal probabilities of measurement to target association, providing fast, accurate estimates. The excellent performance of BP in the particular formulation used may be attributed to the convexity of the underlying free energy which it implicitly optimises. This paper studies multiple scan data association problems, i.e., problems that reason over correspondence between targets and several sets of measurements, which may correspond to different sensors or different time steps. We find that the multiple scan extension of the single scan BP formulation is non-convex and demonstrate the undesirable behaviour that can result. A convex free energy is constructed using the recently proposed fractional free energy (FFE). A convergent, BP-like algorithm is provided for the single scan FFE, and employed in optimising the multiple scan free energy using primal-dual coordinate ascent. Finally, based on a variational interpretation of joint probabilistic data association (JPDA), we develop a sequential variant of the algorithm that is similar to JPDA, but retains consistency constraints from prior scans. The performance of the proposed methods is demonstrated on a bearings only target localisation problem.

💡 Research Summary

The paper tackles the challenging problem of data association across multiple sensor scans in multi‑target tracking by framing it as a variational inference task. While belief propagation (BP) has been shown to work well for single‑scan association because the underlying Bethe free energy is convex, extending the same formulation to multiple scans destroys this convexity, leading to non‑convergence or severely erroneous association probabilities. To overcome this, the authors adopt the recently proposed Fractional Free Energy (FFE), which introduces fractional edge weights (γ∈(0,1]) that soften the mutual‑information terms and restore convexity of the overall free‑energy functional.

First, a BP‑like message‑passing algorithm is derived for the single‑scan FFE. By interpreting the updates as a primal‑dual coordinate ascent on the convex free‑energy, the authors prove global convergence to the unique optimum. Building on this, they construct a multi‑scan model in which each scan has its own FFE term, while consistency constraints from previous scans are enforced via Lagrange multipliers. The resulting optimization remains convex, and a sequential primal‑dual coordinate ascent scheme updates scans one after another, preserving past‑scan marginal distributions.

A novel variational interpretation of Joint Probabilistic Data Association (JPDA) is also presented. Traditional JPDA approximates the joint posterior as a product of marginals, which is non‑convex. By embedding JPDA within the FFE framework, the authors obtain a sequential algorithm that retains JPDA’s marginal‑based philosophy but benefits from the convexity guarantees of FFE.

Experimental validation is performed on a bearings‑only target localisation scenario. The proposed methods are compared against Global Nearest Neighbor (GNN), Multiple Hypothesis Tracking (MHT), standard single‑scan BP, and a naïve non‑convex multi‑scan BP extension. Results show that the convex FFE‑based approaches consistently assign high posterior probability to the true target‑measurement matches, even under high measurement noise and dense target configurations, whereas the non‑convex methods frequently assign vanishing likelihood to the correct solution. The sequential variant further demonstrates that past‑scan consistency can be maintained without a prohibitive computational burden, making the approach suitable for real‑time tracking.

In summary, the work introduces a principled, convex variational formulation for multi‑scan data association, provides a convergent BP‑like solver, and bridges JPDA with modern variational techniques, thereby delivering both theoretical guarantees and practical performance improvements over existing methods.