Authors: Dhakchitha C, Kaviyashree S, Rithika K, Subashini R
Abstract: Maintaining integrity in invigilated examinations is labor-intensive, and human proctors cannot continuously observe every candidate. This paper presents Exam Malpractice Detector, a lightweight, real-time monitoring system built with OpenCV and Python that flags suspicious behavior from a single camera feed. The system combines four inexpensive visual cues: multiple-face detection using a Haar cascade, frame-differencing-based motion analysis, face-position-based "looking away" estimation, and Canny-edge contour analysis for unexpected objects. Detections are mapped to a three-tier status (Normal, Doubt/Review, Confirmed Malpractice) with confidence values. Each alert triggers cooldown-limited, hash-deduplicated, JPEG-compressed evidence capture. Alerts are also projected onto a 5×7 seat grid, where weighted scores produce a color-coded hall heatmap. At session end, a structured PDF report containing session metadata, violation summary, heatmap, event table, and screenshots is generated automatically for staff review. The design keeps a human in the loop: the system flags and documents, and staff decide. Because it needs no GPU and no training data, it can run on commodity hardware.
