In the recent years, more and more modern cars have been equipped with perception capabilities. One of the key applications of such perception systems is the estimation of a risk of collision. This is necessary for both Advanced Driver Assistance Systems and Autonomous Navigation. Most approach for risk estimation propose to detect and track the dynamic objects in the scene. Then the risk is estimated as a Time To Collision (TTC) by projecting the object's trajectory in the future. In this paper, we propose a new grid-based approach for collision risk prediction, based on the Hybrid-Sampling Bayesian Occupancy Filter framework. The idea is to compute an estimation of the TTC for each cell of the grid, instead of reasoning on objects. This strategy avoids to solve the difficult problem of multi-objects detection and tracking and provides a probabilistic estimation of the risk associated to each TTC value. After promising initial results, we propose in this paper to evaluate the relevance of the method for real on-road applications, by using a real-time implementation of our method in an experimental vehicle.