Improving Human Pose Recognition Accuracy using CRF Modeling

Interest in robotics in the domain of manufacturing industry has shown an outstanding growth recently in scenarios where human beings and robots are present simultaneously. Humans and robots often share the same workspace and this poses a lot of threats to the human safety issues [1] e.g. in manufacturing industry, in automobile industry where automobile components are integrated, in medical industry where minimally-invasive-surgery is facilitated and so on. In the proposed approach, segmentation is defined as a classification task and is used for pixelwise object class labeling of human body-parts. Depth measurements from a KINECT RGB-D ceiling sensor are obtained in order to do the pixelwise object class labeling. The ultimate intended use is in the safe human-robot collaboration (SHRC) and interaction (SHRI) domains for challenging domestic and industrial environments. Within this scope, a pairwise conditional random field (CRF) approach is used for labeling. CRF is formulated in terms of an energy minimization (EM) problem while an efficient random decision forest (RDF) is used for classification. In [4], we show how an RDF classifier is used for pixelwise classification of human body-parts using deph data. We found that there exists misclassification of labels assigned to each pixel and that should be minimized for feasible and practical human-robot cooperation. This work builds on top of our previous work Dittrich et al. [4] in order to improve recognizing human body-parts.