An Automated Online Proctoring System Using Pose Estimation and Multi-Model Classification

Synthesis: Arif et al. (2026) built an automated online proctoring system that combines YOLOv11 pose estimation with classical classifiers to flag suspicious behavior in live or recorded webcam video. The pipeline extracts seventeen body landmarks per frame and derives posture deviation, head-movement trend, and upper-body motion features; mutual-information selection kept 30 of them, and single-nearest-neighbor K-Nearest Neighbors was chosen for deployment. On a labeled dataset of examination scenarios the selected model reported 99.56% accuracy, 99.20% precision, 99.60% recall, 99.40% F1, and AUC 0.9957, running from a single webcam through a browser interface with no specialist hardware. The contribution is engineering rather than evaluation: the accuracy figures describe a bespoke labeled scenario set, not live examinations, and the paper does not test whether flagged "suspicious" posture corresponds to actual misconduct or whether the detector treats students with disabilities, cultural differences in gaze, or involuntary movement fairly.

Key Findings

1. Pose estimation replaces the rule-based approach. Rather than thresholding gaze direction or window focus, the system tracks seventeen body landmarks per frame and models deviations in posture, trends in head movement, and unusual upper-body motion, which the authors position against both human invigilation and earlier rule-based monitoring.
2. Feature selection cut the input roughly in half. Mutual-information-based selection retained 30 pose features for classification after extraction, reducing dimensionality before model comparison.
3. Single-nearest-neighbor won the comparison. Standard K-Nearest Neighbors with k = 1 was selected for deployment over the alternatives evaluated on the labeled examination dataset.
4. Reported test performance is near-ceiling. On the held-out test set the deployed model reached 99.56% accuracy, 99.20% precision, 99.60% recall, 99.40% F1, and an AUC of 0.9957.
5. Deployment needs only a webcam and a browser. The interface runs real-time detection on a live stream with no specialist hardware, which the authors present as the route to scalable, low-cost monitoring.
6. Detection is real-time and continuous. The system consumes live or recorded streams frame by frame and provides immediate behavioral feedback rather than post-hoc review of recordings.

What the system does

The pipeline has three stages. Pose estimation runs YOLOv11 over each frame to locate seventeen key body landmarks. Feature engineering converts that landmark stream into quantities describing posture deviation, trend in head movement, and unusual upper-body motion. Classification then maps the feature vector to a suspicious or non-suspicious verdict, with mutual-information selection applied beforehand to keep 30 features and single-nearest-neighbor KNN chosen after a model comparison.

Choosing k = 1 for a proctoring decision is worth pausing on: single-nearest-neighbor classification reproduces the label of whichever training example is closest in feature space, so it carries no smoothing against an unfamiliar posture and inherits every labeling decision made on the training set. The reported near-perfect scores are consistent with a scenario set whose suspicious and non-suspicious cases are visually separable by construction, which is exactly what a controlled recording protocol produces.

What the accuracy figures do and do not establish

A 99.56% accuracy on a held-out split answers one question: given this dataset, this labeling scheme, and this feature set, the classifier reproduces the labels. It does not answer the question an institution actually asks, which is whether a flagged student is more likely than an unflagged one to have cheated.

Three gaps separate the two. First, the dataset is described as "a labeled dataset of examination scenarios" rather than authentic exam sessions, so its base rate and its variety of genuine student behavior are unknown; a detector evaluated on scripted scenarios has not met a student who reads a question aloud, adjusts a lamp, or has a tic. Second, no ground truth about misconduct is reported, so precision and recall are measured against a behavioral proxy rather than against cheating. Third, Remote Proctoring research has long documented that flagging rates and student experience differ across disability status, skin tone, and cultural norms of gaze and posture, and this system does not report any subgroup analysis.

The Privacy exposure is structural rather than incidental: continuous webcam analysis of a student's body in their own room, with the derived features retained or transmitted for review.

Where this sits against the wider evidence

The paper's framing, that rule-based monitoring has failed and that better sensing will fix it, is the same escalation Villanueva (2026) argues against. His audit of 53,915 assessment items found most assessed marks resting on work whose authorship cannot be verified after submission, and his remedy was sufficient evidence of learning at key stages rather than maximum surveillance of every submission. A detector that watches posture more precisely does not change what the assessment can certify.

The evidential standard also matters for how such a system should be read. Published proctoring evaluations that report student outcomes rather than classifier metrics are the ones that bear on Academic Integrity, and those tend to report fear, distraction, and differential burden alongside detection rates. This paper reports none of them, which is a scope limit rather than a flaw in the engineering.

What this means for practice

• Assessment designers. Treat a near-ceiling classifier score as evidence about a labeled dataset, not as evidence that the flagged behavior is misconduct, and ask for the dataset's base rate and labeling protocol before treating a flag as a finding.
• Institutions. Require subgroup reporting before deployment, because posture and gaze conventions vary with disability, culture, and camera position, and a detector calibrated on one population will flag others at a different rate.
• Educational technology developers. Separate the detection claim from the inference claim in the product copy: "the model classified this frame as suspicious" and "this student behaved suspiciously" are different statements, and only the first is supported here.
• Assessment professionals. Compare the surveillance route against evidence-based redesign, since heightened monitoring adds Privacy cost without changing whether the assessment can certify the learning it claims to measure.

Limitations

• The dataset is described only as labeled examination scenarios, with no reported base rate, collection protocol, participant count, or demographic composition, so the accuracy figures cannot be generalized to live invigilated or uninvigilated examinations.
• No ground truth about actual misconduct is reported; precision, recall, and F1 are computed against behavioral labels rather than against cheating, so the system's error rate on the outcome an institution cares about is unknown.
• Single-nearest-neighbor classification with k = 1 gives no robustness to an unfamiliar or boundary posture, and the paper reports no ablation separating the contribution of pose estimation, feature selection, and classifier choice.
• No subgroup or fairness analysis is reported for disability status, skin tone, gender, camera quality, lighting, or cultural norms of gaze, all of which are documented sources of differential flagging in Remote Proctoring systems.
• The evaluation reports no student experience, anxiety, or acceptance measure, and no comparison against human invigilation on the same scenarios, so the claimed improvement over existing practice is not directly measured.
• The version assessed is a journal pre-proof accepted 18 October 2026, so pagination and any post-acceptance corrections may differ from the final article.

Citation

Arif, M. T., Mussa, A. S. M., al Mamun, A., Rahman, M., Rahman, M. M., & Ashiqussalehin, M. (2026). An automated online proctoring system with live suspicious activity detection using YOLOv11 pose estimation and multi-model classification. Computers and Education: Artificial Intelligence, 100688.