Fast Body Posture Estimation using Volumetric Features

Abstract

This paper presents a novel approach to real-time pose recognition using Haar-like features. First, linear discriminant analysis (LDA) is introduced as a powerful new approach to train Haar-like features. The LDA-based method is compared to AdaBoost, and proven to be more efficient and requiring less Haar-like features to successfully complete the pose classification task. The weakened memory requirements with regards to AdaBoost allow for a straightforward extension to a 3D pose detector based on 3D Haar-like features, resulting in a rotation-invariant pose detection system.