An application of structural reliability methods to the road safety domain is proposed in this work. Such methods consider the uncertainties in the model input to deduce a failure probability. Hence, it seems appropriate to handle with the randomness existing in the triplet driver-vehicle-infrastructure. A sensitivity analysis conducted among the model input has shown that four parameters must be modelled by random variables in order to be representative of the uncertainty of the output. This random vector is composed of the total embedded mass, the mass centre position, the lateral position on the lane and the speed perturbation. Each realisation of this vector corresponds to a different trajectory, since the driver model is openloop. The limit between safe and dangerous trajectories is defined by two safety criteria, namely lateral position and lateral acceleration. At this step, reliability indices can be calculated and represent the risk associated with the initial conditions. The application of the proposed method to a real curve shows that the evolution of the lateral acceleration criterion is weakly sensitive to hazard and defines a critical speed for which the threshold of 3 m/s2 is reached. In a sense, this result could be obtained by a deterministic study and so is of little interest here. The comparison between Monte Carlo and Hasofer-Lind index results leads to a satisfying agreement over a useful range of speed for this turn (50-85 km/h). Thus, the number of simulations needed is extremely reduced and more complex models for vehicle dynamics and driver behaviour can be envisaged. The list of safety criteria can also be modified without equipment concerns or changes in the approach. The in situ implementation of this method requires a device measuring initial conditions positioned on the side of the road. Speed and lateral position measurements are easily realisable today using only impedance sensors in the pavement. Difficulties come from the mass and centre of mass position evaluation. On-board measurement may become possible soon if the development of active suspensions makes accessible the vertical forces applied to the wheels. Statistical characteristics of the initial conditions may be identified from the measurement devices accuracy. The associated reliability index may then be calculated or read in a precalculated database, and its value is a decision support tool for alarm triggering since dangerous situations can be deduced by comparison with an acceptable risk. However, the choice of this acceptable risk is a political decision that must result from an expert assessment. Further work will concern the reference trajectory, using data collected from other tasks of the SARI project. Static video sensors, in combination with telemeter, are used to evaluate vehicle trajectories in the real traffic. Hence, position and speed of a great number of common driver can be classified according to their driving style and used to define the reference trajectory. By this way, even hidden effects can be taken into account as, for example, a visibility issue leading drivers to shift sidewards. Another aspect envisaged to reinforce the probabilistic model is to include time-dependent perturbations, particularly in the steer angle. This can be interpreted as the difference between the steer angle that the driver plans to realise and the one actually applied. This involves the notion of stochastic process and requires more complex solution methods.