Multiple Face Tracking using Kalman and Hungarian Algorithm to Reduce Face Recognition Computational Cost

Abstract

Currently, research in face recognition systemsmainly utilized deep learning to achieve high accuracy. Usingdeep learning as the base platform, per frame image processingto detect and recognize faces is computationally expensive,especially for video surveillance systems using large numbers ofmounted cameras simultaneously streaming video data to thesystem. The idea behind this research is that the system does notneed to recognize every occurrence of faces in every frame. Weused MobileNet SSD to detect the face, Kalman filter to predictface location in the next frame when detection fails, andHungarian algorithm to maintain the identity of each face.Based on the result, using our algorithm 87.832 face that mustbe recognized is reduced to only 204 faces, and run at the realtime scenario. This method is proven to be used in surveillancesystems by reducing computational cost.