NUST Institutions Library Catalogue

Analyzing attention enables educators to assess student engagement and enhance
learning experiences. It provides valuable insights for optimizing teaching and
managing classroom behavior. Several techniques have been proposed to analyze
attention and provide feedback to the instructor for effective learning. These include
intrusive and non-intrusive techniques which utilize EEG headsets, eye trackers,
Kinect sensors, cameras, non-verbal cues etc. Intrusive techniques provide accurate
results only for controlled environments prioritizing precise measurements. Moreover,
they cause discomfort to the subjects involved. Whereas non-intrusive techniques
using non-verbal features do not cause any discomfort to the user and can be used in
any environment. However, none of the studies so far have addressed all non-verbal
features simultaneously. This paper presents a multimodal architecture which
integrates all non-verbal features including headpose orientation, body posture
estimation, emotion detection and Eye Aspect Ratio (EAR) calculation to analyze
attention. A deep learning model has been trained on the Facial Expression
Recognition Plus (FERPlus) dataset with 94% accuracy. We used Euler angles to
determine the head pose which includes up, down, left, right and forward directions.
Further EAR is calculated for both eyes using eye key points and Euclidean distance
which shows the opening and closing state of the eyes. Finally estimated the body
pose of the student by training an SVM model & body key points which include
shoulders, elbows, and wrists. The combined result of all these features is displayed in
the form of a graph which reflects the level of attentiveness of the students to the
teacher in real-time. This system can assist the teacher in addressing concerns such as
poor academic performance, disengagement from studies, and high dropout rates
among students.