Understanding dynamic scenes at road intersections is both crucial and challenging for intelligent vehicles. In order to detect potentially dangerous situations, algorithms are needed that can interpret the behaviour of the actors in the scene and predict its likely evolution. The difficulty of this task arises from the large number of possible scenarios. The conventional answer to this issue is to discard vehicle interactions in the manoeuvre prediction process, i.e. to infer the manoeuvre performed by each vehicle from its past and current behaviour, independently from the other vehicles in the scene. In this paper we show how this affects collision risk estimation in very common scenarios, making it unusable in practice for Advanced Driver Assistance Systems (ADAS) applications. As an alternative we propose a probabilistic model for vehicles traversing unsignalised intersections that accounts for the mutual influence between vehicle manoeuvres. The focus is on the utilisation of contextual information (i.e. layout of the intersection, presence of other vehicles and traffic rules) to interpret a vehicle's behaviour. We show how the model can be used for accurate situation and risk assessment.