Human activity classification using a single accelerometer

Abstract

Human activity recognition is an area of growing interest facilitated by the current revolution in body-worn sensors. Activity recognition allows applications to construct activity profiles for each subject which could be used effectively for healthcare and safety applications. Automated human activity recognition systems face several challenges such as number of sensors, sensor precision, gait style differences, and others. This work proposes a machine learning system to automatically recognise human activities based on a single body-worn accelerometer. The in-house collected dataset contains 3D acceleration of 50 subjects performing 10 different activities. The dataset was produced to ensure robustness and prevent subject-biased results. The feature vector is derived from simple statistical features. The proposed method benefits from RGB-to-YIQ colour space transform as kernel to transform the feature vector into more discriminable features. The classification technique is based on an adaptive boosting ensemble classifier. The proposed system shows consistent classification performance up to 95% accuracy among the 50 subjects.