Feasibility study on evaluation of audience's concentration in the classroom with deep convolutional neural networks

In this paper, we developed an estimation system for degree of audience's concentration by estimating individual's behavior with a deep learning approach. Our system firstly detects candidate location of audiences (CLAs) from the movie with Ada-boost classifier composed of Haar-like filters and their integration process. Then, each CLA is investigated to determine the target audience is “concentrated”, “not concentrated” or “no exist” with 5-layered deep convolutional neural networks (DCNN). We used a total of 13 movies of which 3 movies were used for training of DCNN and the remains for evaluation. Our system achieved audience detection performance of precision = 84.8% and recall = 61.8% and estimation accuracy of individual attention as 72.8%.