Object detection on an ordinary laptop: lessons from an exam app
Md Nazmul Hosan ·
How we ran phone and person detection at about 8 to 12 frames per second on student laptops, without sending video anywhere.
Short answer: use a small model (YOLOv8 Nano) through ONNX Runtime, keep a lighter fallback model ready, and smooth detections over time so one odd frame never ends an exam.
For a US exam-prep company we built a desktop app that proctors practice exams on the student's own laptop. The constraint that shaped everything: no webcam video leaves the device.
Choosing the model
We needed to detect people, phones, earbuds, books and watches. YOLOv8 Nano, exported to ONNX and run with ONNX Runtime, gave about 8 to 12 frames per second on an ordinary laptop CPU, which is enough for proctoring. COCO-SSD in TensorFlow.js serves as a fallback on machines where ONNX Runtime struggles.
Faces, gaze and liveness
MediaPipe face mesh gives face presence, head pose and eye landmarks. Liveness combines blink detection, a random blink challenge and small head movements, which rejects photos and replayed video.
False alarms are the real enemy
A student who glances at the ceiling to think is not cheating. Every signal goes through confidence thresholds, minimum durations and consistency checks over several seconds before it becomes an event. Events get a severity score, and only repeated critical events pause a session.
Privacy by design
Reviewers see scored events and screenshots, not a video stream. Biometric processing needs clear consent and retention limits, which we design with each client. We never add emotion detection.
Read the AI Proctoring App case study or see our computer vision service.
Share
Nazmul founded Door Soft in Dhaka and leads system design and delivery. He writes about building AI that works in production for mid-size companies, and about the products Door Soft runs itself.
About the author