Human activity recognition using accelerometers values in different coordinate systems Samad Zabihi * Shahrood University of Technology, Semnan, Iran. Pouyan Nasrollahi School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran. Mohammad Bazmara School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran. *Corresponding author: samadzabihi@outlook.com Keywords Abstract Human activity recognition Intelligent systems Accelerometers Classification Human activity recognition is a technology to recognize activities of one or more subjects from a series of observations on the subject’s body movements. It is an essential technology for service robots and intelligent systems which are in interaction with human beings. This paper proposes a procedure for improving human activity recognition using3-axial accelerometers data. These accelerometers measure acceleration of body motions in three axes, x, y and z. An evaluation on a real-world dataset shows that the quality of features can be improved by transforming accelerometers values to different coordinate systems and extracting features from these transformed data, and then excellent classification results outperforming recent works can be achieved. 1. Introduction Artificial Intelligence is a very broad area and many areas like Image Processing use the expertise of Artificial Intelligence to build systems. Artificial Intelligence provides various methods to be applied in different areas such as really finding soccer talent and goalkeeper quality [1, 2]. Its applications in agriculture [3, 4] or selecting systems of important features [5] and also geographic information systems using fuzzy uncertainty [6]. In modern societies, the number of the elderly who live alone is increasing. In many cases, these people need to be assisted in their daily life. Also, with the development of intelligent systems, these systems are expected to interact well with humans. So both these necessities made it obligatory for intelligent systems to recognize human activities. Nowadays some technologies are developing to solve these problems. Human activity recognition algorithms have improved considerably in recent years. Image processing and use of wearable sensors are two approaches that commonly used for HAR. The image processing approach does not require the use of equipment in the user’s body, but imposes some limitations such as restricting operation to the indoor environments, requiring camera installation in all the rooms, lighting and image quality concerns and, mainly, users’ privacy [7]. The use of wearable sensors minimizes these problems, but requires the user to wear the equipment through extended periods of time. Hence, the use of wearable sensors may lead to inconveniences with battery charges, positioning, and calibration of sensors [8]. A Dynamic Neural Networks (ANNs) approach on physical human action classification implemented in [9] based on dynamic programming model; which makes it similar to our work. Another effort toward action classification has been demonstrated using Evidence Feed Forward Hidden Markov Model (EFF-HMM) in [10]. The EEF-HMM has been used for the gesture recognition by generalizing the movement of people [10][10]. A 3D Human Body-Part Tracking and Action Classification method that builds a Hierarchical Body Model has been proposed using Hierarchical Annealing Particle Filter (H-APF) algorithm in [11]. With the purpose of reducing the use of redundant features in human activity recognition, Mark Hall’s selection algorithm based on correlation was applied in [12]. Also in robotic field, a new method was proposed to classify and choose the best action in humanoid robots using the classification algorithms and machine learning techniques [13]. In this work we present a procedure for extracting efficient features from data of wearable acceleration sensors. By selecting a suitable coordinate system based on the anatomy of human body, recognition rate can be improved without extracting extra feature. The paper is organized as follows: the evaluated dataset is described in Section II. Preprocessing stage will be described in Section III. The quality of features in two of coordinate systems will be evaluated in Section IV. Feature extraction procedure is detailed in Section V and the result of classification with neural network is presented in Section VI. Conclusions will follow in Section VII. 2. Real-World Dataset The In this paper, a public domain HAR dataset have been used. This dataset was provided by [12], and was collected during 8 hours of activities, 2 hours with each one of the 4 subjects: 2 men and 2 women, all adults and healthy. The protocol was to perform each activity separately. The profile of each subject is shown in table 1[12]. Although the number of subjects is small, the amount of data collected is reasonable (2 hours for each subject) and the profile is diverse: women, men, young adults and one Elder. At total it was collected 165,633 samples for the study; the distribution of the samples between the classes is illustrated in figure 1 [9]. World appl. programming, Vol(4), No (12), December, 2014. pp. 237-240 TI Journals World Applied Programming www.tijournals.com ISSN: 2222-2510 Copyright © 2014. All rights reserved for TI Journals.