Tirza Routtenberg

Smart transportation systems rely heavily on accurate state tracking for tasks such as autonomous vehicle navigation, traffic monitoring, and cooperative sensing, often operating in environments with partially known and time-varying dynamics. Classical model-based approaches, such as the expectation-maximization (EM) Kalman filter, require accurate dynamical models and thus may degrade under the complex and non-stationary conditions encountered in real-world transportation scenarios. Recent AI-aided tracking methods, including KalmanNet and its extensions, improve robustness to model mismatch, yet they typically struggle to adapt online to evolving dynamics without labeled data. In this work, we propose EM-KalmanNet, an AI-aided tracking algorithm that integrates EM with KalmanNet through deep unfolding. Our design introduces an unfolded EM-based architecture in which the E-step is implemented via an enhanced RTSNet smoother, while the M-step is rendered differentiable, yielding a compact, interpretable, and trainable machine learning model. We further develop a dedicated training methodology for robust operation under partially specified state-space models with block-wise varying dynamics, representative of changing transportation environments. Numerical results demonstrate that EM-KalmanNet enables accurate tracking under complex time-varying dynamics, making it a promising framework for adaptive state estimation in smart transportation systems.