Today's restraint systems typically include a number of airbags, and a three-point seat belt with load limiter and pretensioner. For the class of real-time controlled restraint systems, the restraint actuator settings are continuously manipulated during the crash. This paper presents a novel control strategy for these systems. The control strategy developed here is based on a combination of model predictive control and reference management, in which a nonlinear device - a reference governor - is added to a primal closed-loop controlled system. This reference governor determines an optimal setpoint in terms of injury reduction andconstraint satisfaction by solving a constrained optimization problem. Prediction of the vehicle motion, required to predict future constraint violation, is included in the design and is based on past crash data using linear regression techniques. Simulation results with MADYMO models show that, with ideal sensors and actuators, a significant reduction (45%) of the peakchest acceleration can be achieved, without prior knowledge of the crash. Furthermore, it is shown that the algorithms are sufficiently fast to be implemented on-line.