Implementation of Jetson Nano Based Face Recognition System

Face recognition has improved a lot with recent advances in deep learning. Face recognition is processed in order of face detection, face alignment, and face recognition. In this paper, We used the K-Face dataset consisting of Korean faces for training the proposed face recognition model, and MTCNN is used for face detection. OpenCV was used for face alignment, Resnet-50 was used for the backbone for face recognition, and the loss function was learned by ArcFace method. Furthermore, In addition, an algorithm to receive KISA's face authentication was implemented, and a face recognition system was implemented in Jetson Nano, an embedded system. For KISA's face recognition authentication, I had to implement it as a DLL, so I used OpenCV. For deep learning learning, Python and TensorFlow were used to infer the learned network for KISA authentication using OpenCV. We improved the speed by using TensorRT for real-time performance in Jetson Nano, an embedded system. A GUI program using Qt5 was developed for behavior testing, and images for testing were implemented to selectively use video files, RTSP, and CSI cameras. Experiments confirm that the implemented system is capable of face detection and recognition at a rate of 10 fps.