Face Detection and Recognition in Video-Streams

Using the Viola Jones face detection algorithm and Active Appearance Models, it is shown that high success rate face-recognition in video-streams is possible with a relatively low processing time. Limitations in the Active Appearance Model Software by Stegmann et al. [15] forced us discard the original thought of doing a recognition based on parameters from a match, using a general model of the human face. Instead we decided to build an Active Appearance Model representing the subject being searched for, and use this model to do a recognition based on statistical thresholds of deviance. Tests have proven very high success rates. Detection rate of faces in the videofootage reached 98,38% with only 1,60% erroneous detections, recognition rates per frame reached 66,67% with 0% erroneous and finally the overall sequence recognitions proved a rate of 88,90% while maintaining 0% erroneous recognitions. The test results clearly indicates that Active Appearance Models are capable of doing high quality face recognitions. Extending the software in order to search for more than one face can easily be done, the computing time however will be doubled whenever the number of Active Appearance Models are being doubled. Had it been possible to do the parameter based recognition, the computing time of the recognition would have remained the same, however recognition of multiple faces would not have any noticeable effect on the computing time.