Knowledge of the state of balance of the moving human is essential for the design of fall detection algorithms and wearable robotic controllers for balance assistance. Still, there is no known general method to make an online estimation of the state of balance of the walking human, with the limited and local sensor measurements that are usually available. Previous work has failed to address such an estimation for human walking rather than standing or biped walking, with only an upper body Inertial Measurement Unit (IMU), and by incorporating the major human balance strategies, to adequately predict the interaction between the human and the robot. In this study, a state estimation technique is introduced: the applied observer is an Additive Unscented Kalman Filter (Additive UKF), and the model consists of a spring-loaded inverted pendulum and articulated upper body, with virtual pivot point (VPP) control and foot placement based on the extrapolated center of mass (XCoM); the Virtual Pendulum Model. The following is described: the dynamic model and observer design, a sensitivity analysis with simulation data, and observer performance with data from a human walking experiment on a treadmill. Proper tuning and limited errors in model parameters, particularly foot contact detection, resulted in promising estimates. With further research toward improved parameter estimation and higher efficiency for online implementation, this method could be useful for the prediction of human movement.