Human action recognition from multi-sensor stream data by genetic programming

Abstract

This paper presents an approach to recognition of human actions such as sitting, standing, walking or running by analysing the data produced by the sensors of a smart phone. The data comes as streams of parallel time series from 21 sensors. We have used genetic programming to evolve detectors for a number of actions and compared the detection accuracy of the evolved detectors with detectors built from the classical machine learning methods including Decision Trees, Naïve Bayes, Nearest Neighbour and Support Vector Machines. The evolved detectors were considerably more accurate. We conclude that the proposed GP method can capture complex interaction of variables in parallel time series without using predefined features. © Springer-Verlag Berlin Heidelberg 2013.