Face Detection and Recognition Methods using Deep Learning in Autonomous Driving

One of the objectives of deep learning is to solve high complex tasks such as perception. In recent years, it has been demonstrated that deep learning can overcome traditional algorithms in image classification as well as object recognition and face recognition tasks. In this paper we are inspecting techniques of deep learning that deals with topical issues in the field of Computer Vision: real-time face detection and face recognition using embedded system and GPU processing on NVidia Tegra X2 (Jetson TX2). In the first part of our work we are proposing a novel experimental research to the problem of face detection and recognition in autonomous driving that use a new deep convolutional neural network model, named FADNet. The architecture model was used on a existing dataset containing more then 13.000 images of 2.000 different faces from different cultures, on which we gained an accuracy of 81.78%, along with an accuracy of 84.45% on a detection dataset containing new 8.600 images. In the final phase of the experimental research we did a real-time test on a dataset of self-acquired video frames from Jetson TX2 embedded system camera, achieving an accuracy of 67.45%, which is a promising result for real-time processing. Last but not least, accuracy and inference time are taken into account by comparing time performance between CPU and GPU implementations.