Pattern recognition-based real-time end point detection specialized for accelerometer signal

Abstract

End point detection is proposed for motion detection by acceleration. Apart from the conventional methods based energy feature normalization in automatic speech recognition and heuristic threshold-based algorithms, supervised learning in pattern recognition is proposed to discriminate a motion state and a non-motion state. Before the algorithm developments in earnest, feasibility and feature selection for the research objectives are mainly studied in this paper. As feature candidates for data representation, we have chosen the absolute value of acceleration, its 1st derivatives, and 2nd derivatives respectively based on correlation coefficient first. Using them, we have formed feature vectors and then transformed 2D or 3D feature vectors into variant vectors with Principle component analysis and Fisher's Linear Discriminant (FLD). Also the sequence of the absolute 1st derivatives with incremental order is critically considered as feature vectors. In addition to the various feature vectors, artificial neural network has been designed to investigate and analyze the feasibility of the proposed algorithm. As a result, it is observed that vectors except for the FLD-transformed doesn't show significant difference and the sequence of the absolute 1st derivatives record comparatively reliable and stable recognition rates regardless of subjects.