A framework for interpreting, modeling and recognizing human body gestures through 3D eigenpostures

In this article we propose a novel system for recognizing human gestures through acquisition and processing of volumetric data sequences. Volumetric sequences are acquired with two different approaches, a multi-camera set-up and a multi-Kinect\(^\mathrm{TM}\) set-up. The recognition based on volumetric representation does not require any skeleton fitting or limb tracking and the system relies on the extraction of robust features directly from the available 3D data. Volumetric shape descriptors are, in fact, invariant with respect to viewpoint and body size; they are designed to provide us with a unique signature for each posture. Hidden Markov Models (HMMs), trained on different gestures, are then used for identifying a set of key postures and classifying their sequences over a set of possible actions. The paper also presents a method for identifying the number of hidden states of the HMMs that describe gestures. Despite its implementation and conceptual simplicity, the number of states that we estimate with this method turns out to match that of other classical approaches such as the Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC). The same approach is also applied in the definition of the Gaussian Mixture for the Hidden states Observations providing us with good results. Extensive tests were performed on a database that we acquired, which is made of ten different actions, each performed by five different actors and in five different ways (different speed and orientation) and on another public database, achieving a 96% correct recognition rate.