Robust Face Recognition System for Extensive Distances in Uncontrolled Environments on Edge Device

Abstract

Traditional CCTV monitoring, reliant on human observers, faces scalability and accuracy challenges in identifying key individuals (dignitaries, faculty) to enable secure access. The rapid advancements in deep learning models have revolutionized the field of face recognition, offering a diverse range of highly effective solutions such as the DeepFace, FaceNet, VGG-Face, and ArcFace. Current face recognition technologies for extensive distances may suffer from accuracy limitations in crowded or uncontrolled environments and may require ongoing manual intervention for optimization and maintenance. An enhanced Multi-Task Cascaded Convolutional Neural Network (MTCNN) algorithm and ArcFace implementation using Resnet-50 is proposed to recognize individuals over medium to long ranges on edge devices such as Jetson Nano. Enhanced MTCNN excels in detecting faces with varying scales and orientations, crucial for recognizing distant faces. ArcFace implementation using ResNet-50, with its deep architecture, learns discriminative features even from low-resolution images, enhancing recognition accuracy over traditional models. For training and testing of the proposed model, manually collected images of individuals under varying illumination and environments and the CelebA dataset consisting of over 200,000 celebrity images, each annotated with various attributes such as hair color, presence of eyeglasses, facial expression, and gender were used.  Compared to the existing works, the proposed MTCNN and ArcFace implementation using the Resnet-50 system demonstrates significantly higher accuracy of 97% with minimal acquisition of dataset to train and scalability, offering a robust and efficient solution for dignitary tracking.