Monitor Attention Of Pupils By Head Pose Estimation And Anomaly Detection

It can be uneasy to stay focused on a speaker.Even teachers sometimes loose the attention of their audience.But,when this audience is composed of young pupils,it is adults' responsibility to make sure that they stay focused on the lesson.Academic success of young people substantially lays on in-class behavior.Looking at the proper spatial element — teacher or board for example — could be regarded as the first and principal step to mental focus.In this thesis,we propose a novel approach to monitor pupils' attention using classic surveillance videos,based on the estimation and classification of “gaze points”.This method raises two main difficulties:· the limited resolution of student' pictures from video surveillance,· and the sorting of acceptable and inappropriate gazes.Indeed,the gaze of each student is deduced from his facial landmarks,which allow us to know his head pose(by solving the PnP problem).But the resolutions of the students' pictures that we study are often below 40x40.To address this challenge,we combine a super-resolution method to a recent landmarks detection method(ensemble of regression trees)in order to more accurately find the facial feature points of the pupils who sit at the last few rows of the classroom.To achieve the second hard task — classification — and detect the children who do not pay attention to the proper spatial area,we consider the distribution of the gaze points in two dimensions.After what we discriminate the anomalous points on a density criterion.Experiments on real videos prove the efficiency and effectiveness of this process.