Designs and Comparisons of Facial Direction Detection Technology with YOLO Based Deep Learning Networks

Abstract

By utilizing the deep learning model to develop the pedestrian facial direction classifier, in this paper, the proposed You only look once (YOLO)-based deep-learning technology is applied to analyze the images captured by camera to identify the facial directions of pedestrians. To enhance the training effect of mirror categories, the selected images are horizontally flipped to expand the datasets. To avoid misclassification of facial directions, the softmax scheme of the original YOLOv2 model is replaced with the logistic classifier used in YOLOv3, and the improved model is called YOLOv2- logistic. By the same parameters on the webcam captured dataset, the experimental results show that the YOLOv2-logistic has the best performances on recall, precision, and mean Average Precision (mAP), which are 85%, 81%, 86.28%, respectively. The YOLOv3 tiny has the second performances, and its recall, precision, and mAP is 81%, 80%, 78.83%, respectively. In the frame per second (fps) test, the YOLOv3 tiny has the second best performance, and it reaches 30 fps on the Xavier platform. Although the facial direction detection performance of the YOLOv3 tiny model is slightly lower than that of the YOLOv2-logistic model, the performance of fps by the YOLOv3 tiny model is more than that by the YOLOv2-logistic model.