Measurement of Human Concentration with Multiple Cameras Kazuhiko Sumi, Koichi Tanaka, and Takashi Matsuyama Graduate School of Informatics, Kyoto University, Kyoto 606–8501, Japan, sumi@vision.kueee.kyoto-u.ac.jp, WWW home page: http://vision.kuee.kyoto-u.ac.jp/ Abstract. We propose a new method to estimate human change of con- centration from multiple camera views of the human. In our method, human state of concentration is observed as self-load, defined as energy injected in a period to keep and manipulate his/her body. If a person is concentrating to a certain task, he/she will brace himself/herself for bet- ter results, and energy consumption will increase. To confirm our idea, we developed a method to calculate self-load from multiple view of the human. We conducted an experiment in which test subjects have dif- ferent level of complexity of task. Self-load of the subjects showed the positive correlation with the complexity of the task. We have convinced that self-load can be used to characterize the concentration of person being observed. 1 Introduction One of the big difference between human-to-human interaction and human-to- computer interaction is timing. We human can measure timing to start con- versation with another human. Measuring timing is a very sophisticated social action, which is not achieved by a computer yet. To realize such a behavior, it is important to have a good sensing systems to observe the human action and his/her internal state. In this research we focus on sensing human internal state, such as concentration, interest and frustration. So far, many studies on human observation were carried out. Most of them are recognizing intentional signal in communication. Such researches include ges- ture recognition, facial expression recognition, lip reading, and voice recognition. Those methods are quite reasonable when the person is already interacting with a computer. On the other hand, if a person is not involved in communication, but is engaging in other personal jobs, the channel of communication is not estab- lished. In such cases, a computer should estimate the human interest, intention and feeling through one-way observation. This will often happen when assisting a human involved in a work, such as driving a car, operating a machine, trav- eling in an unfamiliar places, looking for something, and so on. If the person is doing something in concentration, it is not a good manner to offer help. But if the person is wondering, it might better to support him/her. Thus observing a human without interaction is a challenging subject.