Improved Privacy-Secured Face Direction Estimation Using Point Cloud Data and Deep Learning
R. Takagi, Chinthaka Premachandra · 2024
In recent years, face orientation estimation methods have often relied on capturing facial feature regions, such as the eyes and nose, using the general RGB camera positioned directly in front of the individual. However, these methods encounter challenges when facial feature regions are obscured, such as by a mask, or when they cannot be accurately recognized due to changes in face orientation. Previous studies have attempted to address this issue by introducing depth cameras and exploring potential solutions. Specifically, these previous studies have focused on estimating face direction across seven horizontal classes using 3D point cloud data obtained from depth cameras. However, the impact of mask-wearing scenarios was not sufficiently addressed in prior research. In scenarios such as preventing distracted driving, limiting detection to seven classes and focusing solely on the horizontal direction is inadequate. Broader detection, including vertical orientation, is necessary. To address this, this study explores the use of 3D point cloud data to estimate face orientation across nine classes, incorporating vertical directions, using deep learning techniques to enhance estimation capabilities. Additionally, this study evaluates the estimation accuracy under mask-wearing conditions, considering current societal circumstances.