The problem of traffic state estimation for large-scale urban networks modeled with MFD dynamics is studied here. Given a network partitioned in a number of regions, aggregated traffic dynamics describe the vehicle accumulation in each region, as well as transfer flows to and from neighboring regions. Considering that MFD accumulation-based models have been integrated in perimeter control approaches, this work tackles the real-time estimation problem when limited data is available. An estimation engine is developed according to the Extended Kalman Filter (EKF) theory; it seeks to estimate the real state of the multi-region dynamic system based on traffic sensors' measurements. First, a stochastic model is presented for the dynamics of the process (plant). Then, the EKF estimation scheme is described based on a simpler aggregated model of dynamics and some real-time measurements. Estimation accuracy is investigated through detailed micro-simulation of downtown Barcelona by studying a realistic configuration of real-time measurement availability through loop detector data; however, the developed methodology is generic. The state vector we seek to estimate, as well as the available measurements configuration, can be altered according to the application. The proposed methodology is tested both in macro- and micro-simulation; resulting estimated traffic states (i.e., regional accumulations, demands, and distribution of outflows) are compared to actual ones obtained from the stochastic plant. The developed algorithm can be utilized by closed-loop online urban traffic management strategies to feed the estimated traffic state back to the controller.