Real-Time Face Trajectory Classification Using Distance-Based K-NN and EMD on Edge AI Systems

This paper presents an optimized real-time facial trajectory classification system designed for deployment on edge AI platforms. The system integrates deep learning-based detection, centroid tracking, and a distance-based K-Nearest Neighbor classifier using Earth Mover’s Distance (EMD). Performance optimization strategies including pipeline restructuring and inference acceleration achieved significant latency reduction while maintaining 93% classification accuracy. The architecture is structured for efficient edge deployment scenarios such as embedded AI and surveillance systems.