Tracking people is a common problem for developing an intelligent space. To operate effectively, it is essential that an automatic meeting browser have a solution to what we call the "assignment problem." That is, it must know who said what. It must know the names of the participants and correctly assign the appropriate name to the current speaker. This can be achieved by affixing a lapel microphone to each person, assigning the resulting waveform to that person and typing his or her name into the system before proceeding. For practical and everyday deployment, this simple solution has several drawbacks. First, there is the annoyance of preparation: entering participant's names and wiring everyone up before the meeting begins. Naturally, people would prefer to "just walk in and talk". Secondly, unless the lapel microphones are expensive wireless devices, they restrict movement. As in a normal meeting, participants should be free to get up, move around, distribute papers, walk to the whiteboard, etc.

If freedom of movement is granted, occlusion and poor quality of signals are major challenges for identifying people over time. For example, people might turn their faces away from the camera, other objects might occlude the faces or the faces might be too small for a face recognition system. Similarly the speaker ID module will not work well when a person is not close enough to the nearest microphone. Color appearances of people, however, are more robust in a complex scene. In this research, we use color appearances of people as a way to identify them. The procedure for obtaining the color appearance ID is as follows:
segment people from the background
compute histogram for each person
compute model probability for each person

Related paper:

Multimodal People ID for a Multimedia Meeting Browser
Jie Yang, Xiaojin Zhu, Ralph Gross, John Kominek, Yue Pan, Alex Waibel
acm multimedi 1999