Human physical activity monitoring has received an increasing interest by elders’ caregivers, athletes, physicians, nutritionists, physiotherapists and even people who want to check the daily activity level.
Concerning applications for elderly, and taking into account the actual increasing of aging population and decreasing social and economic conditions for elderly daily care, telecare systems have emerging and have been considered as a solution for some of these problems.
In this project it will be described how to extract human postures, postural transitions and walking patterns from motion data recording with Smartphone built-in accelerometer, particularly a tri-axial sensor. The application of this system is to supervisor elderly or physical active people who are interesting in checking or improve their physical level.
Methods to monitor activities of daily living, as standing, sitting, lying, walking and climbing stairs, were proposed based on a dataset composed of ten 60-70 years subjects, five male and five female, who carried a Smartphone placed longitudinally on their waist. The threshold-based approach implemented was capable of discriminating between static and dynamic activities. Static activities refer to situations when the user is static in a posture, as standing, sitting or lying; dynamic activities refer to activities that involve movement of the user, as walking, climbing stairs and transitions between postures (sit-to-stand movements). Within static activities, the angle between the user initial position and the orientation during each movement was used to differentiate the user postural orientation. The analysis of walking patterns were conducted in the frequency domain using Fast Fourier Transform and analysing the peak in the correspondent spectrum with higher amplitude, which corresponds to the step rate.
The problem of discriminating activities can be treated as a classification problem using techniques of machine learning.
Using a public dataset provided by SmartLab during the ESANN competition, it was possible to study the more suitable metrics to extract from acceleration signals in order to train and test a classifier. Using a decision tree classifier, which was implemented with J48 algorithm from Weka, it was possible to achieve 86% accuracy and a 14% classifier error for the train and test datasets provided in the competition.
The results obtained suggest that accelerometer sensors could be used for accurate physical activity detection and strategies to implement the classification algorithms in Android environment should be implemented.
For any additional information regarding this project, please contact us using the inquiries form.