Action detection for concentration degree estimation of TV viewer

Abstract

In recent years, the demand to an internal state of the person estimate technology increases. The non-contact type technique to estimate internal state of in particular is expected in application to purchasing action analysis and human-robot interactive. In this suggestion, I think it to become available as useful information in CM or improvement of the program contents production by understanding a view state of the TV viewer. Unlike daily action recognition, I classify pose of TV viewer by learning joint coordinate of the spine with a Support Vector Machine and detect action of TV viewer by identifying sub-interval with the maximum score of classifier learned with joint coordinates, joint angles and object information around the hand of the upper body because it is thought that they are necessary to estimate a concentration degree of TV viewer.