Face & Eye Detection — Offline attendance prototype
Role: Lead engineer · Stack: C++ · OpenCV · Haar cascades
Problem
Build a privacy-preserving, offline attendance capture that runs on commodity laptops without cloud dependency. Constraints: low CPU, limited memory, and varying indoor lighting.
Solution & architecture
Single-process pipeline: webcam capture → CLAHE preprocessing → Haar-based face detection → eye detection → temporal consensus filter → CSV logger. Design prioritized deterministic CPU usage and predictable latency. CLI for export and basic reporting.
Trade-offs
Haar cascades were chosen over DNN models to avoid GPU requirements and heavy dependencies. This lowered raw accuracy but ensured immediate deployability and stable performance on low-end hardware.
Testing & metrics
Field-tested across multiple sessions, tuned temporal window to reduce false positives. Achieved >85% detection in standard indoor conditions after tuning.
Outcome
Delivered a prototype producing CSV attendance logs. Validated product hypothesis without cloud costs and reduced administrative effort for the client.