In this work, an efficient and accurate face recognition system based on edge processing using GPUs was completely developed. A complete pipeline that contains a sequence of processing steps, including pre-processing, face feature extraction, and matching, is proposed. For processing steps, lightweight deep neural models were developed and optimized so that they could be computationally accelerated on an embedded hardware of Nvidia’s Jetson Nano. Besides the core processing pipeline, a database, as well as a user application server were also developed to fully meet the requirements of readily commercialized applications. The experimental evaluation results show that our system has a very high accuracy based on the BLUFR benchmark, with a precision of 98.642%. Also, the system is very computationally efficient, as the computing time to recognize an ID in a dataset of 1171IDs with 10141 images on the Jetson Nano is only 165ms. For the critical case, the system can process 4 camera streams and simultaneously recognize a maximum of 40 IDs within a computing time of 458ms for each ID. With its high-speed and accuracy characteristics, the developed system has a high potential for practical applications.
Keyword
Face recognition, deep learning, GPUs, edge processing.