EDUTRACK: A CHECKPOINT-BASED FACE-RECOGNITION ATTENDANCE SYSTEM FOR UZBEK HIGHER EDUCATION

Аннотация

Manual attendance in higher education consumes 5–10 minutes of every lecture, scales poorly to large classrooms, and is easily defeated by proxy sign-ins and mid-lecture departures. Prior face-recognition attendance systems address the sign-in cost but overwhelmingly report only laboratory accuracy on small, controlled datasets and do not defend against attendance fraud that occurs during a session. We present EduTrack, a production-grade classroom attendance platform deployed for one semester at Inha University in Tashkent, Uzbekistan. EduTrack combines a Spring Boot orchestration layer, a Python/ONNX recognition engine (YuNet detector + ArcFace R50 512-dimensional embeddings), and a PostgreSQL/pgvector similarity store, achieving a measured P95 end-to-end recognition latency of 47 ms per face on commodity CPU hardware. We contribute (i) a four-window checkpoint-grading scheme that samples attendance across a lecture rather than at a single moment, cutting undetected “sign-and-leave” events by an estimated 93% relative to single-snapshot baselines; (ii) an identity-stabilization rule requiring five consistent matches before recording a recognition, which reduces false-accept rate to an estimated 0.4% at a 0.70 cosine threshold; and (iii) the first reported semester-scale deployment of such a system on a Central Asian student population—a demographic under-represented in face-embedding training corpora. Across three classrooms (N = 168 enrolled students, 12 weeks, 1,240 sessions), EduTrack matched teacher-provided ground truth with 96.2% accuracy, reduced per-lecture attendance overhead from an average of 6.8 minutes to under 10 seconds, and received a mean usability score of 4.3/5 from surveyed teachers (n = 14) and 3.9/5 from students (n = 112), with privacy concerns emerging as the principal adoption barrier.