Existing trackers based on Poisson measurement process often struggle with efficiency and accuracy in large-scale tracking under heavy clutter. To overcome this, we introduce PiVoT, a scalable, robust multi-object tracker capable of efficiently detecting and tracking a large, varying number of objects, along with their shapes, existence probabilities, and measurement rates, even in heavy clutter. PiVoT employs a novel two-stage variational inference routine to achieve inference tractability and closed-form, parallelisable updates. Efficiency is further enhanced by early identification and removal of ineffective birth objects and designing highly simplified, much faster, yet equivalent variational updates. Additionally, PiVoT inherently offers efficient clutter-robust clustering, an innovation that can also enhance existing trackers that depend on supplementary clustering techniques. Experiments demonstrate PiVoT's clear accuracy and efficiency gains over existing methods, while also highlighting its ability to track a thousand closely spaced objects in under a second on a standard laptop without gating.

In recent years, state-space models for highly manoeuvrable objects have been proposed based on non-Gaussian, continuous time, jump-based Lévy processes, the so-called Lévy state-space model [1]–[4]. In these models, the standard Brownian motion driving process for continuous time processes is replaced with a heavy-tailed non-Gaussian alternative. This retains all the flexibility of its Gaussian counterpart in terms of possible dynamical model structures and operations with irregular time stamps or heterogeneous data sources. These models aim to operate in areas such as surveillance of irregularly moving drones or people, and tracking wildlife or biological data. Implementation is relatively straightforward since the Kalman filters of the Brownian motion case can be replaced in the nonGaussian case by mixtures of Kalman filters within a marginalised particle filtering framework [5]. While the Stone Soup tracking software environment includes both Kalman filtering and generic particle filtering, it does not currently allow the combination of these tasks within a marginalised particle filtering framework. We discuss the significant challenges involved in incorporating these models and algorithms into Stone Soup, and present initial simulation results for the new software.

This paper investigates the task of tracking multiple objects in clutter under a distributed multi-sensor network with time-varying connectivity. Designed with the same objective as the centralised variational multi-object tracker, the proposed method achieves optimal decentralised fusion in performance with local processing and communication with only neighboring sensors. A key innovation is the decentralised construction of a locally maximised evidence lower bound, which greatly reduces the information required for communication. Our decentralised natural gradient descent variational multi-object tracker, enhanced with the gradient tracking strategy and natural gradients that adjusts the direction of traditional gradients to the steepest, shows rapid convergence. Our results verify that the proposed method is empirically equivalent to the centralised fusion in tracking accuracy, surpasses suboptimal fusion techniques with comparable costs, and achieves much lower communication overhead than the consensus-based variational multi-object tracker.

This paper addresses the joint detection and tracking of an unknown and time-varying number of targets in clutter. Here we formulate the tracking task in a variable-dimension state space, under which the reversible jump sequential Markov chain Monte Carlo sampling methods can be utilised to online estimate the target number, their kinematic states, and the association variables. In particular, a fast Rao-Blackwellisation scheme is devised to improve the tracking accuracy and sampling efficiency for linear Gaussian models. Based on the nonhomogeneous Poisson process measurement model, the developed tracker enjoys a partially parallel sampling structure, thereby being able to efficiently tackle the data association under massive measurements and clutter. The simulation results demonstrate that the developed tracker exhibits superior tracking performance in comparison to existing trackers in both accuracy and computational efficiency when tracking multiple targets under heavy clutter.

Category-level object pose estimation involves estimating the 6D pose and the 3D metric size of objects from predetermined categories. While recent approaches take categorical shape prior information as reference to improve pose estimation accuracy, the single-stage network design and training manner lead to sub-optimal performance since there are two distinct tasks in the pipeline. In this paper, the advantage of two-stage pipeline over single-stage design is discussed. To this end, we propose a two-stage deformation-and-registration pipeline called DR-Pose, which consists of completion-aided deformation stage and scaled registration stage. The first stage uses a point cloud completion method to generate unseen parts of target object, guiding subsequent deformation on the shape prior. In the second stage, a novel registration network is designed to extract pose-sensitive features and predict the representation of object partial point cloud in canonical space based on the deformation results from the first stage. DR-Pose produces superior results to the state-of-the-art shape prior-based methods on both CAMERA25 and REAL275 benchmarks. Codes are available at https://github.com/Zray26/DR-Pose.git.

Coordinate ascent variational inference (CAVI) is a popular approximate inference method; however, it relies on a mean-field assumption that can lead to large estimation errors for highly correlated variables. In this paper, we propose a conditionally factorized variational family with an adjustable conditional structure and derive the corresponding coordinate ascent algorithm for optimization. The algorithm is termed Conditionally factorized Variational Bayes (CVB) and implemented with importance sampling. We show that by choosing a finer conditional structure, our algorithm can be guaranteed to achieve a better variational lower bound, thus providing a flexible trade-off between computational cost and inference accuracy. The validity of the method is demonstrated in a simple posterior computation task.

In this paper we present multidimensional α-stable state-space models for object tracking, expressed in continuous time as Lévy processes. In contrast with the conventional Gaussian models, these heavy-tailed α-stable models are more likely to exhibit extreme noise values, thus showing the capability for modeling of erratic manoeuvring behaviour. Despite the potential benefits, such models are usually highly intractable for inference and therefore have not yet been widely adopted in the tracking field. Here the models are represented in a conditionally Gaussian series form, so that the marginal (Rao-Blackwellised) particle filter can be employed to perform tracking and smoothing very efficiently. As the result, the simulation tracks present some sharp manoeuvres, owing to the heavy-tailed property, and experiments demonstrate improved performance on an intent inference problem from automotive UI with highly perturbed pointing data.

This letter presents an alternative, more consistent, construction for bridging distributions, which enables inferring the destination of a tracked object from the available partial sensory observations. Two algorithms are then introduced to sequentially estimate the probability of all possible endpoints within a generic Bayesian framework. They capture the influence of intended destination on the object's motion via suitably adapted stochastic models. Whilst the bridging approach has low training requirements, the proposed formulation can lead to more efficient predictors, e.g. around 65% less computations for certain models. Synthetic and real data is used to illustrate the effectiveness of the introduced algorithms.