The motivation is simple and familiar to anyone who has lectured to a hall of eighty students. Instructors judge attention by scanning faces, watching note-taking, body posture and hand-raising, but in large, densely packed rooms this manual assessment is neither scalable nor continuous. Disengagement often goes unnoticed until poor exam results reveal it, weeks after the fact. Automated alternatives exist, but most stream raw high-definition classroom video to cloud servers for deep learning inference, creating privacy risks, compliance headaches under laws such as India’s Digital Personal Data Protection Act of 2023, and dependence on expensive infrastructure and reliable connectivity that many institutions simply do not have.

The researchers’ answer is a privacy-by-design pipeline that runs entirely at the edge. A camera attached to the Raspberry Pi 5 captures a classroom image every sixty seconds throughout a lecture session. Each frame is resized in local memory and passed to RetinaFace, a dense multi-scale face detection network that returns bounding boxes, confidence scores and five-point facial landmarks for every visible student. The landmarks feed an affine alignment step that corrects head tilt, rotation and scale, producing normalized face crops. Those crops then flow into DeepFace, a lightweight facial attribute framework that assigns each face a probability distribution across seven emotion categories: happy, sad, neutral, angry, surprise, fear and disgust.

The heart of the system is a weighted aggregation scheme that converts the class-wide emotion distribution into a single, interpretable attentiveness score between 0.0 and 1.0. Each emotion carries a manually specified attentiveness multiplier ranging from 1.0 down to 0.3, reflecting how strongly that expression correlates with engagement in established classroom mapping models. The weighted counts of detected emotions are normalized by the number of verified faces in the frame, yielding a per-frame score. A running cumulative score tracks the longitudinal engagement pattern across the whole session, and a rule-based classifier maps the continuous value into qualitative labels of low, medium or high attentiveness using fixed thresholds. In one reported example, a frame from a dense classroom scored 0.58, or medium attentiveness, while another scored 0.70, categorized as high.

Crucially, nothing identifiable survives the process. Raw frames exist only in device memory during inference and are automatically deleted afterward. No facial images, embeddings or biometric templates are written to permanent storage or transmitted over the network at any stage. The only records kept locally in a CSV file, and optionally synchronized to a cloud dashboard, are non-identifiable fields: a timestamp, lecture identifier, detected person count, dominant classroom emotion and the attentiveness score. Any synchronization occurs over a TLS-secured channel, and institutions can choose to run the system fully offline with no cloud component at all. The researchers designed the threat model around the risk of centralized biometric data storage, and by keeping all visual inference local, they eliminate the attack surface associated with network interception and server-side breaches entirely.

The engineering choices also make the system fault tolerant. Session directories are created automatically when a lecture begins, a JSON status file tracks whether the system is active or idle to conserve compute during empty rooms, and local-first logging means the pipeline keeps running even when internet connectivity drops. This matters for schools and colleges where reliable broadband cannot be assumed. The cloud dashboard, when used, displays a real-time gauge of instantaneous and session-wide attentiveness alongside trend analytics, giving instructors non-intrusive feedback on the collective state of the room rather than a per-student surveillance record.

The empirical validation took place in two real university classrooms with very different characteristics: one with high seating density and uniform artificial lighting, another with a wider layout and mixed natural and artificial light. Across eight sessions lasting roughly fifty to fifty-five minutes each, with student numbers ranging from sixty to ninety, the system held up under the messy conditions of live teaching, including occlusion between closely seated students and variation in camera angle and illumination. In qualitative checks, the detector reliably localized faces across front, middle and back rows in both rooms.

The quantitative results are striking for hardware this modest. RetinaFace achieved 95.67 percent face detection accuracy, comfortably beating the 89.46 percent of a three-layer convolutional neural network baseline the authors implemented for comparison. DeepFace reached 93.46 percent accuracy in seven-class emotion classification, against just 78.42 percent for the lightweight baseline. Most importantly for real-time use, the complete pipeline, spanning face detection, emotion recognition, aggregation and score computation, averaged 62 milliseconds of inference latency per frame on the Raspberry Pi 5. That figure demonstrates that high-capacity pretrained backbones can deliver both accuracy and acceptable speed on a resource-constrained edge device, at a fraction of the cost of the GPU servers that cloud-centric engagement systems typically require.

The study distinguishes itself carefully from adjacent work. Prior Raspberry Pi deployments have focused on face-recognition attendance with proxy prevention, while other edge-AI systems track per-student engagement trajectories or run in-browser engagement ratios for online learning. None of the compared systems produced an aggregate class-level attentiveness score for physical classrooms. The proposed framework is explicitly a class-level analytics tool rather than an identity verification or attendance system: it reports how engaged the room is as a collective, not who is paying attention. The authors also note an ablation study confirming that the performance gains come from the choice of pretrained backbones, evaluated under identical data, preprocessing and hardware conditions, rather than from other pipeline factors.

Ethical safeguards were built into the study from the start. Consent was obtained from every participating student and instructor, the work was approved at the institutional level, and all classroom images published in the paper were blurred or anonymized. The authors argue that their four-tier design, spanning capture, edge processing, logging and optional cloud visualization, aligns with the data minimization, purpose limitation and retention requirements of modern privacy legislation. Because no mechanism exists to reconstruct individual identity from the stored data, the system offers a template for how engagement analytics could be deployed without turning classrooms into biometric surveillance zones.

Limitations remain, and the researchers are candid about them. The emotion weights and classification thresholds are manually specified design choices rather than learned parameters, making the attentiveness score an interpretable first approximation of engagement rather than a validated psychological construct. Emotion alone is an imperfect proxy for attention, and the current version ignores eye gaze, posture and speech activity. The evaluation covered two classrooms at a single institution, so generalization across room geometries, cultures and lighting regimes is untested. Future work, the authors say, will integrate complementary behavioral cues, explore adaptive and personalized scoring, and push the deep learning models toward greater energy efficiency on edge hardware.

Even so, the demonstration is significant. It shows that a computer costing under a hundred dollars can run industrial-grade computer vision on a room full of faces, distill that signal into a single meaningful number, and do so without ever letting a photograph leave the building. As debates over AI surveillance in education intensify worldwide, this study offers a compelling middle path: continuous, objective feedback for teachers, delivered at negligible cost, with student privacy protected not as an afterthought but as an architectural principle.

Subject of Research: Real-time classroom attentiveness monitoring using edge-AI facial emotion recognition on a Raspberry Pi 5

Article Title: Edge-AI enabled real-time classroom attentiveness monitoring via facial emotion recognition on Raspberry Pi

Article References: Davda, B., Vahora, S., Maltare, N., & Patel, H. (2026). Edge-AI enabled real-time classroom attentiveness monitoring via facial emotion recognition on Raspberry Pi. Frontiers in Imaging, 5, Article 1972869. https://doi.org/10.3389/fimag.2026.1972869

Image Credits: AI Generated

DOI: 10.3389/fimag.2026.1972869

Keywords: edge AI, Raspberry Pi, facial emotion recognition, classroom attentiveness, RetinaFace, DeepFace, privacy-preserving, deep learning, computer vision, student engagement, real-time inference, educational technology