ANN for human pose estimation in low resolution depth images

摘要

The paper presents an approach to localize human body joints in 3D coordinates based on a single low resolution depth image. First a framework to generate a database of 80k realistic depth images from a 3D body model is described. Then data preprocessing and normalization procedure, and DNN and MLP artificial neural networks architectures and training are presented. The robustness against camera distance and image noise is analysed. Localization accuracy for each joint is reported and application for low resolution and large distance pose estimation is proposed. A very fast regression on body joints locations in 3D space is achieved, even in case of sensor noise, large distance and reaching off the screen.