This paper presents an ultra-fast detection algorithm for locating faces in grey scale videos. We first use motion detection to reduce the working area and find the approximate position of the head. Then a morphology-based technique is applied in this area to detect eye-analogue and lips-analogue regions. Next, the resulting components are used to search for potential facial features. Finally we select from the candidate triplets, the one that best represents a real face, calculating a fitness which takes into account things such as the symmetry and the proximity with the extrapolated position of the face. In order to achieve the maximal speed-up, we use the vector parallelism provided by the SIMD (Simple Instruction Multiple Data) extensions, available on most mainstream processors. The final program runs 65 times faster than the real-time. Experiments demonstrate that the success rate for single face videos reaches 85% in good conditions and can go down to 60% in harder cases. This approach can be useful in many applications, where the detection rate is not as important as the computation time, such as video face identification, or human-computer visual interfaces.