Smart Attendance System

Manual attendance is repetitive, slow and vulnerable to proxy attendance. The project explored whether face recognition could automate identity verification while staying aware of the limits and risks of biometric identification. The engineering question was not simply whether a face can be recognised, but whether an automated pipeline can work under realistic conditions, and what happens when it does not.

System design

Camera input, frame processing, face detection, face preprocessing, feature extraction, identity matching, a confidence threshold, then an attendance decision. Confident matches become a timestamped attendance event; everything else becomes unknown or goes to human review. Repeat detections of the same person are ignored.

Method

A sequential computer-vision pipeline: capture frames, detect faces, extract and preprocess face regions, generate and compare facial representations, calculate identity similarity, apply a recognition threshold, record attendance only when the decision is confident enough, and flag uncertain cases instead of forcing a match. It was tested across variations in lighting, facial pose, recognition confidence and matching threshold.

The resulting pipeline demonstrated real-time face detection, identity matching and attendance recording. Performance was sensitive to environmental conditions: recognition was strongest in favourable lighting with a frontal pose and degraded as lighting and pose became more challenging. That reinforced testing a vision system against its specific operating conditions instead of relying on one aggregate figure. No accuracy numbers are quoted here because the original test results are not reproduced on this page.

What it does not show

The original Python implementation is not embedded here. The interactive demonstration on this page is a separate browser implementation: one mode is a simulation with drawn people whose score penalties are illustrative, and the other runs a different face-detection and recognition model in your browser. Neither reports the original system's accuracy, and false-match and bias evaluation are not part of the demo.

Lessons

Lighting and pose change the quality of the facial representation available. Thresholds are engineering decisions with trade-offs. A system that works in ideal conditions can still fail in deployment. And a useful deployment has to answer when it fails, how confident each decision is, what happens when it is wrong, what data it processes, and who is affected.

Responsible AI

A structured ethical review covered privacy and biometric data, consent, false recognition, false rejection, bias, surveillance and human oversight, treated as part of the engineering process. Low-confidence decisions are reviewable instead of being treated as truth.