A Markov blanket feature selection approach for HMMs in human motion recognition application

Abstract

Human motion recognition is a hot topic in the human-robot interface field. This paper investigates human motion recognition based on hidden Markov models (HMMs) using Kinect data. Kinect provides skeletal data consisting of 3D body joints and is inexpensive and convenient. We extract features from the Cartesian coordinates for human body joints for HMM learning and recognition. To reduce the feature dimensions and improve the accuracy of motion recognition, a method for determining the optimal feature subset of HMMs is required. Feature selection methods for static learning mechanism are widely used, but few methods have been applied to sequential data models such as HMMs. Here we propose a novel Markov blanket method for HMM feature selection that is based on a dynamic Bayesian network (DBN) structure learning. In the learned DBN structure, the Markov blanket of human motion label nodes should be the minimal/optimal feature subset for the HMMs. The proposed method is applied to the MSR Action3D data set. Results show that the proposed method yields better recognition accuracy than traditional feature selection methods.