Advancing multi-target tracking through realistic state-dependent modeling for practical applications

<p dir="ltr">With the rapid growth of automation in areas such as advanced driver assistance systems (ADAS), air traffic collision avoidance, and unmanned aerial vehicle (UAV) navigation, reliable perception of dynamic environments has become critical. Multi-target tracking (MTT), which estimates both the number of targets and their states from noisy measurements, lies at the heart of these applications. While automation brings convenience, safety must not be compromised, making accuracy in state estimation essential.</p><p><br></p><p dir="ltr">Traditional MTT methods such as the joint probabilistic data association filter (JPDAF) and multiple hypothesis tracking (MHT) address the problem through measurement-to-track association but suffer from high computational costs in dense or cluttered environments. More recent approaches based on finite set statistics (FISST), including the probability hypothesis density (PHD), Gaussian mixture PHD (GM-PHD), and multi-Bernoulli (MB) filters, avoid explicit association and offer scalable alternatives. However, their effectiveness in real-world scenarios is often limited by simplifying assumptions such as constant target survival and detection probabilities, and birth rate, or by high communication demands in distributed networks.</p><p><br></p><p dir="ltr">This dissertation develops a set of application-driven extensions to the GM-PHD and MB filters to address these limitations. First, a state-dependent GM-PHD filter is introduced, in which target survival and detection probabilities, and birth rate vary with the target state and environment. By incorporating road geometry and occlusion information, this method improves robustness to missed detections and constrained vehicle motion in road traffic surveillance. Second, a multiple-model GM-PHD filter with state-dependent probabilities is developed to track maneuvering targets along structured trajectories such as roads, sea lanes, or air routes. Hybrid system modeling is used to define mode transitions based on state-dependent guard conditions, capturing realistic target while maintaining tractable recursion. Third, motivated by UAV surveillance applications, a state-dependent multi-target multi-Bernoulli filter with a closed-form Gaussian mixture implementation is developed. By explicitly modeling target survival and detection probabilities, and birth rate as state-dependent functions, the proposed method captures UAV launch, landing, and occlusion behaviors while ensuring consistency with the true cardinality distribution. Finally, motivated by the need for scalability in sensor networks, a distributed GM-PHD algorithm under communication constraints is proposed. This method employs weighted arithmetic average fusion with a probabilistic sampling strategy, providing convergence guarantees while significantly reducing communication cost and mitigating false-positive amplification.</p><p><br></p><p dir="ltr">The effectiveness of the proposed methods is demonstrated through simulation studies across road, maritime, aerial, and distributed sensing scenarios. Results show improved estimation of both target states and cardinality compared to baseline GM-PHD and MB-family filters, while maintaining computational efficiency. Collectively, this work bridges advances in FISST-based filtering with application-driven requirements, offering robust, scalable, and practical solutions for modern multi-target tracking.</p>