The safety assessment of autonomous driving system is a major challenge in the automotive industry and the role of simulation in development and testing of autonomous vehicles has become predominant to significantly reduce the hundreds of millions of miles required to demonstrate the safety performance of such systems. In this paper, a novel methodology is presented to assess automated vehicles safety performance based on a multifactorial analysis of severity indicators in the vicinity of the under test self-driving car. The set of severity indicators includes commonly used time intervals (Inter Vehicular Time, Time to Collision, Time to Steer, ...), distance-based indicators, traffic congestion indicators and a newly developed indicator relying on overlapping geometrical regions. Unsupervised clustering techniques are then used to investigate the correlations, dependence among the whole set of indicators. To address the problem of combining these heterogeneous quantities to derive a global measure of dangerousness for a given scenario, appropriate scaling is performed and various aggregation methods are tested against cut-in, cutout and cut-through scenarios.