摘要: An algorithm for speaker's lip motion detection is presented, based on the processing of a colour video sequence face under natural lighting conditions and without any particular make-up. It intended applications in speech recognition, videoconferencing or synthesis animation. The statistical approach using Markov Random Field (MRF) modelling, with spatiotemporal neighbourhood pixels image sequence. Two kinds observations are used : temporal difference between successive images (motion information) purity red hue current past (spatial information about location). field hidden labels, relevant detection, obtained by energy minimisation proves to be robust (shadows). This label extract qualitative (mouth opening closing) but also quantitative measuring some geometrical features (horizontal vertical spacing) directly field.