This paper deals with face detection and tracking by computer vision for multimedia applications. Contrary to current techniques that are based on huge learning databases and complex algorithms to get generic face models (e.g. active appearance models), the proposed method handles simple contextual knowledge representative of the application background thanks to a quick supervised initialization. The transferable belief model is used to counteract the incompleteness of the prior model due first to a lack of exhaustiveness of the learning stage and secondly to the subjectivity of the task of face segmentation. The algorithm contains two main steps: detection and tracking. In the detection phase, an evidential face model is estimated by merging basic beliefs elaborated from Viola and Jones face detector and from a skin colour detector, for the assignment of mass functions. These functions are computed as the merging of sources in a specific nonlinear colour space. In order to deal with colour information dependence in the fusion process, the Denoeux cautious rule is used. The pignistic probabilities stemming from the face model guarantee the compatibility between the belief framework and the probabilistic framework. They are the entries of a bootstrap particle filter which yields face tracking at video rate. We show that the proper tuning of the evidential model parameters improves the tracking performance in real-time. Quantitative evaluation of the proposed method gives a detection rate reaching 80%, comparable to what can be found in the literature. However the proposed method requires only a weak initialization.