Sequential data feature selection for human motion recognition via Markov blanket

Human motion recognition is a hot topic in the field of human-machine interface research, where human motion is often represented in time sequential sensor data. This paper investigates human motion recognition based on feature-selected sequential Kinect skeleton data. We extract features from the Cartesian coordinates of human body joints for machine learning and recognition. As there are errors associated with the sensor, in addition to other uncertain factors, human motion sequential sensor data usually includes some irrelative and error features. To improve the recognition rate, an effective method is to reduce the amount of irrelative and error features from original sequential data. Feature selection methods for static situations, such as photo images, are widely used. However, very few investigations in the literature discuss this with regards to sequential data models, such as HMM (Hidden Markov Model), CRF (Conditional Random Field), DBN (Dynamic Bayesian Network), and so on. Here, we propose a novel method which combines a Markov blanket with the wrapper method for sequential data feature selection. The proposed algorithm is assessed using four sets of open human motion data and two types of learners (HMM and DBN), and the results show that it yields better recognition accuracy than traditional methods and non-feature selection models.