Estimating body pose of infants in depth images using random ferns

In recent years, many systems for motion analysis of infants have been developed which either use markers or lack 3D information. We propose a system that can be trained fast and flexibly to fit the requirements of markerless 3D movement analysis of infants. Random Ferns are used as an efficient and robust alternative to Random Forests to find the 3D positions of body joints in single depth images. The training time is reduced by several orders of magnitude compared to the Kinect approach using a similar amount of data. Our system is trained in 9 hours on a 32 core workstation opposed to 24 hours on a 1000 core cluster, achieving comparable accuracy to the Kinect SDK on a publicly available pose estimation benchmark dataset containing adults. On manually annotated recordings of an infant, we obtain an average distance error over all joints of 41 mm. Building on the proposed approach, we aim to develop an automated, unintrusive, cheap and objective system for the early detection of infantile movement disorders like cerebral palsy using 3D motion analysis techniques.