Article Preview
TopIntroduction
Smartphones are carried by almost everyone in their daily lives, becoming more powerful every day. Since the invention of the first smartphone (Sarwar & Soomro, 2013), many facilities have been introduced in the daily lives of its users, including sending emails, consulting the remote bank account or take pictures and videos. Twenty years later, the number of smartphones sold in the world has surpassed the number of laptops. It owners reached 5 billion and researchers are still trying in this context to implement hard and soft devices that improve the lives of people(Sivakumaran & Iacopino, 2018). In recent years, smartphones have been used for Human Activity Recognition (HAR), because they are readily equipped with several sensors useful for activity recognition, such as motion and location sensors.
Physical activity recognition using smartphone sensors has enabled the scientific community to develop novel applications. HAR is widely used in the fields of sports (Minetto, 2015), healthcare and assisted living (Candás, Peláez, López, Fernández, Álvarez, & Díaz, 2014), industry, etc. We are interested in this work to the HAR system and consider them for the benefit of the elderly and people with health problems, for example, patients with obesity, osteoarthritis, or heart disease. These people are often required to follow a well-defined daily routine as part of their medical care. Indeed, the use of one device that the patient already possesses is an important thing to reduce the costs of medical care bills, and the frequency of his medical visits. Thus, a doctor can follow the patient’s profile remotely and frequently improves his profile.
Although the idea of a portable system, lightweight, cheap, not bulky and able to establish the user’s profile is very promising, the task of recognizing human actions remains problematic. This is due, on the one hand, to the redundant information in the data, which makes difficult the extraction of the physical postures and then leads to misinterpret the recognition of human activities. On the other hand, another main difficulty comes from the great intra-class variations of a human activity. Indeed, great variations of style can be observed in the reproduction of the same action according to gender, size, and habits, or in the context in which the action is performed. In addition, inter-class ambiguity must be considered for similar actions such as ‘run slowly’ (jogging) and ‘walk’ which differ only by their speed. Thus, a reliable human activity recognition system must have a large capacity for generalization, taking into account intra-class variations while being able to discriminate actions having a small inter-class variation.
All of today’s smartphones are equipped with different available sensors. It is an energy-limited device for obtaining an optimal recognition of the human activity. The Accelerometer has received the most attention in the activity recognition research. However, other sensors, including the gyroscope, and magnetometer have been combined to improve the activity recognition performance. With the use of these sensors, it could be possible to exploit the movement speed, the rotation, the geographical location, etc. Moreover, users carry their phones throughout the day while driving, walking, and running, etc. The sensors can therefore provide information on different possible physical activities throughout the period of the performed activity.
Human activity recognition requires running classification algorithms, originating from statistical machine learning tools (Dougherty, 2012). Mostly, supervised or semi-supervised learning techniques are used and such techniques rely on labeled data, i.e., associated with a specific activity class. HAR using smartphone data is a basic multi-variables time series, for which the task is to analyze, process and classify those contiguous portions of sensor data streams that cover various activities of interest. Hence, user-independent training and activity recognition are required to foster the use of human activity recognition systems where the system can use the training data offline from other users in classifying the activities of a new subject online.