We propose a Bayesian model and inference method for tracking multiple extended objects with a latent directional group structure. Groups of objects are defined by a common directional mode, i.e., the objects in a group move in the same direction and with the same speed up to random fluctuations. The number of directional groups, the parameters defining the groups, and the group affiliations of the objects are unknown; they are learned online, simultaneously with the tracking process. Our online learning method is based on a novel directional motion model with a Dirichlet process nonparametric prior for the group parameters, and it uses a Gibbs sampler for joint parameter estimation and object clustering. The Gibbs sampler is combined with a recently proposed extended multiobject tracking method using the belief propagation algorithm. Simulation results show that exploitation of the objects' latent directional group structure using the proposed method significantly improves the performance of extended multiobject tracking.

Most multiobject tracking methods rely on a statistical model that involves unknown parameters. Here, we propose a Bayesian method for class-aided online learning of model parameters within extended multiobject tracking. We address the case where the extended objects belong to unknown object classes defined by unknown values of the model parameters. The proposed method learns the number of object classes, the class parameters, and the objects’ class affiliations simultaneously with the tracking process, and the learned class and parameter information is leveraged for improved tracking. This is enabled by a parameter-dependent state-space model for extended multiobject tracking that incorporates a Dirichlet process prior, and by a related Gibbs sampler for online learning. Our simulation results demonstrate substantial gains in tracking performance due to class-aided online parameter learning.

We consider a distributed labeled multi-Bernoulli (LMB) filter that uses the generalized covariance intersection technique for fusing the local LMB distributions. A critical aspect of such filters is to correctly associate labeled Bernoulli components describing the same object at different sensors. Here, we improve on previously proposed association schemes by introducing a probabilistic framework and algorithm for object (label) association. Instead of enforcing a hard association, we propose to compute association probabilities and use them in the fusion of the LMB posterior distributions. To develop our probabilistic label association scheme, we first derive a formulation of the fused multiobject distribution that involves a label association distribution. We then show that approximating the label association distribution by the product of its marginals results in a fused multiobject distribution that is again of LMB type. An efficient LMB fusion algorithm is finally obtained by using a belief propagation scheme for fast approximate marginalization and a Gaussian approximation. Simulation results demonstrate that the resulting distributed LMB filter outperforms a state-of-the-art method using hard label association.

The Bernoulli filter (BF) is a Bayes-optimal method for target tracking when the target can be present or absent in unknown time intervals and the measurements are affected by clutter and missed detections. We propose a distributed particle-based multisensor BF algorithm that approximates the centralized multisensor BF for arbitrary nonlinear and non-Gaussian system models. Our distributed algorithm uses a new extension of the likelihood consensus (LC) scheme that accounts for both target presence and absence and includes an adaptive pruning of the LC expansion coefficients. Simulation results for a heterogeneous sensor network with significant noise and clutter show that the performance of our algorithm is close to that of the centralized multisensor BF.