This work is a part of the RaViOLi project and is about multi-objects tracking for driver assistance. In this project, the experimental car is fitted with : a radar, a lidar and a stereovision system. A list of detected objects is provided by each sensor including their coordinates given in distance and angle. This information is treated with the belief functions theory which change data in belief masses. Then, a tracking algorithm, also based on belief functions theory, is used to localize again the detected objects in the new objects list provided by the sensors. This association is made by comparing the old objects list to the new one between each sample time, and when an object match in the two lists, it is considered to be the same. An existing tracking algorithm, developed by M. Rombaut, was used and it was modified to fit the constraints of project RaViOLi. A comparison between these two methods of association is shown, as well as their limits and the advantages of the modifications made. Finally, a prediction step is used to predict objects positions at two different times. First, a t + dt second prediction is computed, where dt corresponds to the sampling time of the sensor. This prediction is used as new data in the association step to improve the tracking. Then, a prediction at t + n seconds is done, where n depends of the car speed. Dangerous cars can thus be extracted and an alert can be sent to the driver if necessary. Synthetics and real data are used to test the robustness of the algorithm in several situations.