Vision-Based Model for Continuous Monitoring and Evaluation of Multiple Digital Screen Exposure

Pre-pandemic, online learning had already been established as an alternative mode of delivery in some schools. Yet with the COVID-19 onset, the majority of instruction has been migrated online, with most schools maintaining a hybrid setup at present. This setup requires end devices to access learning materials and classes; hence, a significant increase in screen exposure due to academic-related activities on top of leisure has been recorded within this period. Research suggests that lower physical and psychological well-being can be directly associated with excessive screen usage. Existing solutions in the form of self-report methods, software trackers, and sensors introduce gaps such as parents miscalculating their children's screen time on questionnaires; limited distance and viewing angle of sensors; and inability to track screen time of multiple devices at once. To overcome these limitations, this study models a vision-based system capable of detecting and classifying multiple devices, and quantifying screen time per classification. Datasets of different device types were divided into stationary and handheld classes, and underwent YOLOv5s detection training. Following successful detection, a screen time counter was programmed to accumulate time per class. Presented herein is the system having detection performance of 0.958 and 0.882 mean average precision for Intersection over Union (IoU) thresholds 50 and 50–95, respectively. Meanwhile, the screen time monitoring achieved a minimal average percentage error of 1.14%. With a maximum coverage of 60 centimeters and two meters for handheld and stationary, respectively, the system proves to be an efficient supplement to existing solutions for screen usage monitoring.