16 September 2024 in IEEE Transactions on Aerospace and Electronic Systems
This paper introduces an innovative approach for resolvable group target tracking (RGTT) through the integration of classical Bayesian filtering and sophisticated deep learning techniques. Numerous extant RGTT methods oversimplify target motion models and interactions within groups, limiting their applicability and compromising tracking performance. Deep learning methodologies are recurrently leveraged to overcome these challenges owing to their adeptness in delineating intricate mappings. We develop a neural network tailored for RGTT, incorporating a neural ordinary differential equation (Neural ODE) to articulate the motion model of targets. Additionally, it predominantly integrates recurrent neural network (RNN) and graph neural network (GNN) modules. Similar to multi-target tracking tasks, the RNN module is enlisted here to model nonlinear dynamic systems. The distinction lies in the domain of resolvable group targets, where an additional reliance on graph theory becomes imperative for estimating group structures. Consequently, this paper introduces a GNN module into the deep learning network to learn the interaction relationships among targets within the same group. Furthermore, the seamless fusion of this deep network with the labeled multi-Bernoulli (LMB) filter augments the pragmatic viability of the deep learning algorithm Simulation results elucidate the efficacy of the proposed algorithm.
2 December 2025 in IEEE Sensors Journal
This paper integrates two data-driven learnable modules into the conventional distributed multi-sensor resolvable group target tracking (RGTT) framework based on the labeled multi-Bernoulli (LMB) filter to improve tracking performance. First, in RGTT scenarios with densely distributed targets, the LMB filter is prone to track fragmentation and label switching, which reduce accuracy and complicate multi-sensor data association. While some studies have attempted to address this, their methods typically overlook the group structure and are often offline and computationally intensive. To overcome this, we propose a track optimization module based on attention mechanism that alleviates these issues efficiently in an online setting. Second, we replace traditional data association algorithms, which rely on fixed similarity metrics, with a track matching module based on message passing network (MPN). This module leverages historical track data and learns to infer latent similarities across sensors, significantly improving matching accuracy. Simulation results demonstrate the effectiveness of the proposed modules.