Abstract

As surveillance cameras are everywhere, there’s a strong need for systems that can automatically spot and track suspects without constant human monitoring. Looking at CCTV manually is slow and prone to error. In this paper, we introduce FaceTrace AI that works in two modes. SCRFD is used for face detection, and ArcFace ResNet-50 is adopted to generate 512-dimensional embeddings. Known suspects are recognized by a linear SVM trained on L2-normalized embeddings, and unknown suspects are detected using cosine similarity with user-provided images. The model is tested based on 619 images from three subjects with 1 FPS thresholds of 0.68 and 0.55. Powered by the Dockerized FastAPI backend, FaceTrace AI records important attributes and generates a PDF with timestamping, confidence scores, and annotated frames—achieving precision at commercial scale.