Real-Time Face Tracking and Vector Databases for Scalable Face Recognition

Real-time face recognition systems increasingly require flexibility in deployment across diverse hardware, from resource-constrained edge devices to high-performance servers, while supporting scalable identity retrieval as databases grow from thousands to millions of entries. We present ScaleEdgeFace, a modular and parallel face-recognition framework for reproducible benchmarking and deployment on resource-constrained and high-performance hardware. The system decouples capture, detection, tracking, embedding, and vector search into concurrent threads linked by lock-free queues, enabling independent scaling and fair cross-backend comparisons. We integrate multiple detectors (FaceBoxes, MediaPipe, YOLOv8) and recognizers (FaceNet, Inception-ResNet) yielding 128-D embeddings with standardized alignment. Retrieval uses pluggable vector databases: NumPy for local operation and Pinecone for million-scale galleries. Accelerated with ONNX Runtime and TensorRT (FP16/INT8), the best configuration (MediaPipe+Norfair) achieves 350.7 FPS (max 412.4) on RTX 2070 Super and 57.7 FPS (max 90.2) on Jetson Nano, while maintaining 94% verification accuracy on LFW and 92% on VGGFace2. ScaleEdgeFace demonstrates scalable real-time performance across hardware tiers, from edge-only to cloud-connected deployments with vector-database- backed retrieval.