Activity Recognition Using Multiple Inertial Measurement Units

Objectives: This paper addresses the design of an ambulatory monitoring system based on a set of wearable, wireless inertial measurement units able to perform activity recognition for healthy individuals and Parkinson's disease patients, as well as analyze and assess the severity of levodopa induced dyskinesia.Material and methods: The monitoring system is composed of six Shimmer3 modules placed at different positions of the individual's body. Both healthy individuals and one patient performed a protocol of simple daily life activities while wearing the Shimmer3 modules. As an initial step, validity of the monitoring system in identifying healthy individuals' activities is assessed. Data corresponding to the activities was separated and features in both time and frequency domains were extracted. Multiple factor analysis was used to evaluate and infer the relationships between the different module positions. A method of feature selection was implemented to determine the most important features, positions and sensors included in the different modules. The classification of activities was done using a KNN classifier.Results: Promising results were obtained in classifying the activities of healthy individuals, with a global accuracy of 77.6%. However, certain adaptation is required for the application on Parkinson's disease patients.Conclusion: While activity recognition for healthy individuals using this system was successful, further evaluation of the contribution of each module needs to be done in order to determine optimal module positions. To validate the obtained results on Parkinson's disease patients, a larger study based on more patient acquisitions is envisioned.