Nostril detection for robust mouth tracking

Within an Audio-Visual Speech Recognition (AVSR) framework an important process is video feature extraction. Several methods are available, but all of them require mouth region extraction. To achieve this, a semi-automatic system based on nostril detection is presented. The system is designed to work on ordinary frontal videos and to be able to recover brief nostril occlusion. Using the nostril position a motion compensated Accumulated Difference Image (ADI) is generated. This ADI is less noisy than the non-compensated one, and this leads to better mouth region tracking. Results show that the ADI stage has good reliability, whereas the nostril detection stage may be further improved.