BrightPath AI: A Privacy-Preserving Edge-AI Framework for Early Warning of Classroom Engagement, Safety, and Student Well-Being

Abstract

Schools increasingly use digital tools to support learning, yet continuous observation of individual engagement and safety remains difficult in large classrooms, while continuous cloud video analytics raises substantial privacy concerns. This paper presents BrightPath AI, an edge-first educational early-warning framework that combines open-set face identity routing, classroom-behavior classification, facial-expression category detection, violence detection, a questionnaire-based well-being risk screen, and role-specific mobile reporting. Raw classroom video is processed on a local edge unit and discarded after inference; downstream services receive only derived metadata and alerts. The framework was evaluated using the project datasets and implementation logs supplied for five analytical modules. At a cosine-similarity threshold of 0.5, the face-verification module produced 0 false acceptances among 46,548 impostor comparisons and 27 false rejections among 862 genuine comparisons, corresponding to 99.94% pairwise accuracy, 96.87% recall, and an observed false-acceptance rate of 0%. The classroom-behavior model achieved 98.69% test accuracy across seven classes. The YOLOv8m facial-expression detector obtained mAP@0.50 of 0.877 and mAP@0.50:0.95 of 0.697. The MobileNet-BiLSTM violence detector achieved 90% accuracy on the reported 40-video test subset, with no false-positive violence alerts in that subset and 110.55 ms inference time per 15-frame sequence. The binary well-being screening classifier achieved 95% accuracy and 0.92 recall for the needs-support class. The scientific contribution is the integration of heterogeneous school-safety and well-being signals within a privacy-minimizing, human-in-the-loop architecture, together with explicit operational boundaries that prevent facial-expression or questionnaire outputs from being treated as autonomous clinical diagnoses or disciplinary decisions.