Face Direction Estimation based on Mediapipe Landmarks

Abstract

Many studies had been presented to estimate face direction that relied on training images. A small portion of these studies relied on mathematical calculations with poor results. This paper proposed implementing of trigonometric functions to estimate face direction by using detected landmarks. These landmarks had been detected using the 3D model of (Mideapipe) method. Among there detected landmarks, several points are selected as predicted landmarks. The angles among extended lines from these selected landmarks have been measured to calculate the directions of the human head. The experimental results showed that proposed method achieved high performance with very low error ratio (0.055) for Yaw angle, (0.046) for Pith angle and (0.025) for Roll angle. The difference was approximately ±0.5 degree between real and predicted measures.