Head pose estimation from 2D/3D images

Abstract

The purpose of the current work is to develop, test and compare different methods for the detection of the angles of the head. To achieve good head estimation this project has been divided in two stages. In the first stage, different face detection and tracking techniques have been implemented, as well as a face point clouds constructor. In the second stage, 4 different techniques for head pose estimation have been implemented and tested using the “ICT 3D HeadPose Database [75, 76]” to evaluate their behaviors and make a comparison between their results. After a presentation and analysis of the main state-of the-art head pose estimation algorithms, we present our four selected methods based on the use of 3D images (PCA, cylinder fitting using RANSAC, cylinder fitting doing an exhaustive search of the cylinder axis and ICP). These methods are completely automatics, so they do not need any training. In order to evaluate the results and identify potential error sources, two different analyses have been done: (1) a global analysis for each method tested over the whole database and (2) a specific analysis using different techniques over some subjects of the database. From the results of these analyses, a comparison between the four methods is made.