Human Activity Recognition via the Features of Labeled Depth Body Parts

This paper presents a work on labeled depth body parts based human activity recognition. In this work, we label depth silhouettes for various specific body parts via trained random forests. From the labeled body parts, the centroid of each body part is computed, resulting in 23 centroids from each depth silhouette. Then from the centroids in 3D, we compute motion parameters (i.e., a set of magnitude and directional angle features). Finally, Hidden Markov Models are trained with these features and used to recognize six daily human activities. Our results show the mean recognition rate of 97.16% over the six human activities whereas a conventional HAR approach achieved only 79.50%. Our system should be useful as a smart HAR system for smart homes.