Extreme Learning Machine for Linear Dynamical Systems Classification: Application to Human Activity Recognition

This paper proposes a Extreme Learning Machine (ELM) recognition framework for human activities using essential dynamic characteristics of the activity. Raw activity time series are collected from inertial sensors embedded in smart phone.We model each activity sequence with a collection of linear dynamical system (LDS) models, each LDS model describing a small patch of the sequence. A codebook is formed using the K-medoids clustering algorithm and a Bag-of-Systems (BoS) is developed to represent the activity time series. Then use ELM to classify them. Great advantages of this method are that complicated statistical feature design procedure is avoided and the LDSs can well capture the dynamics of the activity. Our experiment validation on public dataset shows promising results.