Real-Time Student Engagement Monitoring on Edge Devices: Deep Learning Meets Efficiency and Privacy

Since the advent of COVID-19, educational institutions worldwide have increasingly adopted online learning where challenges in maintaining student engagement have been observed. Consequently, innovative solutions for monitoring and improving engagement in virtual environments have been explored. This paper presents the design and implementation of an affordable, efficient device equipped with a camera that infers student engagement in real time using Deep Learning techniques. The system utilizes a convolutional neural network to analyze visual data and classify student behavior into three categories; focusing on paper, focusing on the screen, and being distracted. To ensure compatibility with resource-constrained edge platforms, the trained model was quantized using TensorFlow Lite (TFLite), enabling efficient deployment across a variety of devices, ranging from low-power microcontrollers like the Sony Spresence and OpenMV cam to more powerful GPU and TPU-based edge platforms such as NVIDIA Jetson Orin and Google Coral. We also established a baseline on the Raspberry Pi 5, optimizing for performance and resource constraints. A key focus of the work presented here was safeguarding student privacy. All data processing occurs locally on the device, ensuring that no sensitive information is transmitted externally. The quantized model enables real-time performance without compromising privacy, as all engagement data is securely analyzed on the edge device; only the student's engagement state is transmitted across potentially insecure networks, while the sensitive video feed remains confined to the student's device. A comprehensive comparison between the quantized, non-quantized, and quantization-aware models was conducted to assess the impact of quantization on accuracy and inference speed. In addition, trade-offs between power consumption and performance across different devices were evaluated. These findings demonstrate a balanced approach to achieving computational efficiency while ensuring real-time student engagement monitoring. By leveraging these cost-effective technologies, our approach offers an accessible and scalable realtime student engagement monitoring tool and addresses privacy challenges posed by remote education, even beyond the pandemic.