Study of Feature Extraction Techniques for Sensor Data Classification

Abstract

Human activity recognition is a rapidly growing area in healthcare systems. The applications include fall detection, ambiguous activity, dangerous behavior, etc. It has become one of the important requirements for the elderly or neurological disorder patients. The devices included are accelerometer and gyroscope, which generate large amounts of data. Accuracy of classification algorithms for this data is highly dependent upon extraction and selection of data features. This research study has extracted time domain features, based on statistical functions as well as rotational features around three axes. Gyroscope data features are also used to enhance accuracy of accelerometer data. Three popular classification techniques are used to classify the accelerometer dataset into activity categories. Binary classification (run -1 / walk-0) is considered. The results have shown SVM and LDA when used with rotation and gyroscope data gives the highest accuracy of 92.0% whereas FDA shows 84% accuracy.