A new skin colour estimation method based on change detection and cluster analysis

Skin colour is a useful and robust cue for face and hand detection and tracking and it has been widely exploited in many different applications such as human motion analysis or image and video retrieval. The presented work is a result of the joint Fraunhofer/Max-Planck research project AVATecH (Advancing Video/Audio Technology in Humanities Research), which deals with automatic annotation of large video corpora in humanities research for psycholinguistics. One of the tasks in this project is the development of a robust skin colour estimation algorithm necessary for the realization of tools for hands and head tracking. Because of the peculiarities of the large video corpora, there are no skin colour models applicable to detect the correct skin colour in each video. Therefore an algorithm was developed that estimates skin colour independently for each video and more important, it works reliably without the need of a training set. The algorithm uses both a change detection tool, to select the most suitable frames for skin colour estimation, and an iterative clustering algorithm to select the range in the YUV domain that best represents skin colour