FedTrack: Federated Learning for Localization and Velocity Estimation of Moving Targets

We introduceFedTrack: a novel federated learning (FL)-inspired algorithm for distributed estimation of a moving target's location and velocity. Unlike traditional approaches, which rely on either centralized processing or independent local estimates,FedTrackenables collaborative sensing across distributed devices. This enables efficient joint estimation from distributed samples. The joint estimator is a function of the Doppler shift in the received signal which is a function of target's location and velocity. Since the joint estimator requires a large number of samples from the target's trajectory, we formulate the joint maximum likelihood (ML) estimation problem as a distributed optimization task, where devices locally collect sufficient statistics and compute individual log-likelihoods. By treating these local log-likelihoods as loss functions,FedTrackeffectively distributes the sensing task, allowing devices to iteratively refine a global estimate of the target's position and velocity. This communication-efficient method closely approximates the joint ML estimate while significantly reducing reliance on a central coordinator. Numerical results demonstrate that FedTrack's estimation accuracy closely tracks the Cramér-Rao Bound (CRB) and outperform conventional baselines like one-shot distributed sensing and closely follow centralized sensing.