DESIGN AND IMPLEMENTATION OF AN EXAMINATION ANTI-CHEATING SYSTEM IN COMPUTER LABORATORIES BASED ON LIGHTWEIGHT COMPUTER VISION

Abstract

With the deep advancement of educational informatization, computer-based examinations in university computer laboratories have become a critical approach for evaluating students' practical skills. However, traditional centralized video invigilation systems impose severe computation and network bandwidth strains, making them impractical for legacy hardware deployments. To address these bottlenecks, this paper designs and implements a lightweight, multi-dimensional behavior auditing invigilation system based on decentralized edge computing. By offloading image capturing, posture estimation, and decision-making workflows entirely onto the localized client side, the proposed system restricts network transmissions to low-frequency, structured text logs, fundamentally eliminating the risk of campus network congestion caused by concurrent high-definition video streaming. Specifically, on the algorithmic layer, the system leverages the native features of the 44 homogeneous transformation matrix from the WebAssembly-accelerated MediaPipe Tasks-Vision framework. It performs inverse kinematics for head yaw and pitch via basic scalar operations in the JavaScript runtime, completely bypassing heavy dependencies such as OpenCV.js. Concurrently, a geometric mathematical model utilizing Euclidean distance ratios of iris and pupil landmarks is established for eye gaze tracking. Furthermore, a dual-layer cooperative anti-false-alarm mechanism integrating spatial directional branching and a temporal sliding window filter is proposed, which smoothly filters out transient physiological noises and benign gestures. Experimental results on a simulated low-end host demonstrate that the system achieves an exceptionally low single-frame computation latency of 0.62 ms and restricts the total memory footprint to 146.5 MB. Under steady-state operations, it reduces the false positive rate (FPR) of environmental and physiological noise to 1.3% while maintaining a high true positive rate (TPR) of 96.3% for actual cheating behaviors. The system exhibits outstanding efficiency, robust anti-interference capability, and high deployment feasibility for inclusive application on legacy educational equipment.