AI-Powered CCTV Exam Monitoring System
Real-time AI proctoring for exams: unauthorised items flagged, identity verified, invigilators given a live dashboard — 0.92 mAP at under 15 ms per frame.
- Python
- YOLOv8
- OpenCV
- PyTorch
- DeepStream
- TensorRT
- Triton
- Flask
- React
- DeepFace
- Docker
- ByteTrack / DeepSORT
- Kafka / MQTT
- C++ / CUDA
What it is: an edge-deployed, real-time AI proctoring system that monitors camera feeds (USB / RTSP / file) to detect unauthorised items such as phones, with multi-camera tracking, biometric verification and a live operator dashboard.
What I did: led a 4-person team and built a hardware-aware pipeline orchestrator that selects between a full NVIDIA Metropolis pipeline (DeepStream, TensorRT FP16/INT8 YOLOv8, Triton Inference Server over gRPC), a Hybrid TensorRT pipeline, and a CPU-only legacy PyTorch pipeline — with ByteTrack / DeepSORT tracking, Kafka / MQTT event streaming, and C++/CUDA performance extensions. I also added offline VLM-generated incident reports, an AES-256 encrypted rolling evidence cache, a benchmarking suite, and Docker + GitHub Actions CI deployment.
Results, measured on the capstone: 0.92 mAP at IoU 0.5, under 15 ms per frame at 30 FPS, and 98.2% session-scoped biometric accuracy.
Technology: Python, YOLOv8, OpenCV, PyTorch, DeepStream, TensorRT, Triton, Flask, React, DeepFace, Docker, ByteTrack / DeepSORT, Kafka / MQTT, C++/CUDA.