Proposal of a Concentration State Estimation System Integrating Gaze Estimation and Non-contact Pulse Estimation

Abstract

We propose a system that estimates the user's state of 'concentration/distraction/drowsiness' in real time, using only camera footage from smartphones or similar devices as input. Based on face detection via MediaPipe, we introduce a front-facing baseline (neutral) from the Yaw/Pitch estimated by L2CS-Net to calculate relative gaze direction. We then detect 'looking away' or 'darting glances' using gaze velocity (measured on two scales: short-term and long-term). Furthermore, it estimates BPM from pixel value fluctuations in the Region of Interest (ROI), treating a 7% increase above the baseline as tension and a 20% decrease as a sign of drowsiness. These are integrated via a serial gate, incorporating a holding time to suppress frame-level fluctuations. By visualising and logging the state label alongside supporting indicators, it enables self-management support during learning or work.