Driver assistant systems in general, and specially collision avoidance systems are more and more installed in recent vehicles because of their high potential in reducing the number road accidents. Indeed, those systems are designed to assist the driver or even to take its place when the risk of collision is very important. This thesis deals with the main challenges in the development of collision avoidance systems. In order to react in a convenient way, the system must, first, build a faithful representation of the environment of the ego-vehicle. Perception is made by means of exteroceptive sensors that detect objects and measure different parameters, depending on their measurement principle. The fusion of individual sensor data allows obtaining a global knowledge that is more accurate, more certain and more varied. This research work makes a deep exploration of high level multisensor, multimodal, multitarget tracking methods. The proposed approaches are evaluated and validated on real driving data and also on simulated scenarios. Then, the observed scene is continuously analyzed in order to evaluate the risk of collision on the ego-vehicle. The thesis proposes methods of vehicle trajectory prediction and methods to calculate the probability of collision at different prediction times. This allows defining different levels of alert to the driver. an automotive scenarion simulator is used to test and validate the proposed scene analysis approaches. Finally, when the risk of collision reaches a defined critical value, the system must compute a collision avoidance trajectory that will be automatically followed. The main approaches of trajectory planning have been revisited et one has chosen according to the context of driver assistant system.
