Identification and classification of motor activities using inertial sensors data

Activity of daily living (ADL) recognition has great importance in the field of rehabilitation, physical monitoring and ubiquitous computing. Recognition of ADL is essential because of the connection between physical inactivity and common health problems, like osteoporosis, cardiovascular disease, diabetes, obesity. This study is motivated by the fact that it is important to monitor the activities of a person in daily routines, so as to associate the day-by-day motor performance with the recommendations given by the physicians. A set of aerobic activities (walking, stairs walking, running, sitting, standing) which are considered useful to promote the well-being of a person are used to design the activity recognition system. Inertial measurement units including accelerometer and gyroscope represent a promising technology in long term physical activity monitoring. This work aims to design a physical activity recognition system intended to classify motor activities from an inertial sensor, which can be profitably used in real-time applications. To achieve this goal, the system is designed and evaluated on different parameters of the algorithm: pre-processing steps involved in signal processing, segmentation of the signal to minimize delays associated with further processing, determination of the best feature set for classification, and training of the classification scheme based on both subject-dependent and subject-independent validation to maximize recognition accuracy. It is believed that the accuracy of systems able to recognize daily living activities in real time heavily depends on the processing steps for signal segmentation. This study presented a modified event-based segmentation algorithm that introduces a much reduced temporal delay between activity occurrence and its detection and recognition. The system is capable of distinguishing those activities which are considered hard to differentiate in the previous studies, such as walking, stairs ascending and stairs descending. The presented scheme is not only capable to classify these activities, but it can also be helpful in ambulatory gait analysis, where on-time processing may be needed, and spaces for motion capture systems are not at hand. Finally, the availability of the inertial data flows coming from mobile phones and smart bracelets makes it possible to include the detection and recognition algorithms presented in this thesis into these commodities, thus expanding their use also for daily living activity long term monitoring for fitness, active ageing, and †œactive growing†� applications.