# Students' posture sequence estimation using spatio-temporal constraints

In recent years, lectures have been analyzed as a part of Faculty Development. There are many analyses focused on the behavior of students. However, since they record the students' behavior manually, observers must bear heavy tasks. Thus, it is useful to estimate the students' behavior by computer automatically. In order to implement such system, we need a standard description scheme to describe students' behavior in classrooms.
In this paper, we propose more primitive criterion than behavior, which is called posture sequence. The goal of this study is to estimate students' posture sequence from videos recorded in classes. A posture sequence is a time-series of a student's postures during a lecture and a posture of a student is described by a set of his head, body trunk and hand states, which we call the body segments states. A behavior can be dened as a specic part of the posture sequence.
Since students' body segments often overlap in the observed videos, detection of students' body segments is difficult. To cope with this problem, we introduce the spatio-temporal constraints of body segments and postures. We call these constraints 'posture model' and 'transition model'. The posture model is the spatial constraints between body segments and applied in estimating the posture of each moment. The transition model is the temporal constraints between postures. The posture sequence estimation process holds evaluation value of each posture in some interval, and propagates evaluation value through the interval. Using these constraints, we gradually revise the posture sequence during the
interval and output appropriate estimation results.
In the experiment, we apply our proposed method to a real lecture, and show that our method can improve the accuracy of posture sequence estimation. We also present a preliminary example of behavior classication using our method.