Fast and efficient integration of human upper-body detection and orientation estimation in RGB-D video

Automatic and accurate human upper-body detection and orientation estimation have great practical value in several computer vision applications. Most previous works on human upper-body orientation estimation assume that the human upper-body region is already detected and aligned. However, this is not the case in many real-world scenarios. Additional human detector is essential which is usually much slower than orientation estimator. In this paper, we propose a novel approach to single-frame human upper-body detection and orientation estimation in RGB-D images. Unlike previous works which address human detection and orientation estimation separately with different models, our method involves a random forest framework to approach both in a unified model. The whole algorithm, including human detection and orientation estimation, achieves 40fps in 640p resolution on a single laptop CPU without any GPU acceleration. The experimental results show the effectiveness of the proposed method.