Human activity recognition using smartphone's sensors and machine learning
Enrique García Ceja
Summary or description
In this thesis human activity recognition is performed from data gathered through the sensors of smartphones. Human activity recognition is an important task for ambient intelligence systems. Being able to recognize the state of a person can provide us with valuable information that can be used as input for other systems. For example, in health care, fall detection can be used to alert the medical staff in case of an accident; in security, abnormal behavior can be detected and thus used to prevent a burglary or other criminal activities. This work focuses on physical and daily living activities. The first type refers to activities that can be inferred by just analyzing them for a few seconds (2-10 seconds) and are independent of the situation, e.g., walking, resting, running, etc. The second type refers to activities that are composed of a collection of physical activities, e.g., working, shopping, exercising, etc. We also used information gathered from Wi-Fi access points to contextualize the physical activities in order to have a better understanding of the user’s situation. In this case, context is defined as the additional relevant information that gives sense to the user’s actions. The activity recognition task was stated as a classification problem and Machine Learning methods (supervised and unsupervised) were used to perform the classification. For physical activities recognition, K-nearest neighbors, Decision tree and Naïve Bayes methods were used, achieving overall accuracies of 89.33%, 87.33% and 93.33%, respectively. Simple methods and computationally efficient features were looked for given that the implementation is intended for small devices with scarce resources. Once the physical activity is recognized, the next step is to put it in context. Since every Wi-Fi access point has a unique identifier, it is possible to use this information along with the received signal strength to locate the user. For example, if the user is walking and the Wi-Fi access points’ identifiers correspond to the ones installed in a library, then it could be inferred that the user is looking for a book. Since the access points’ identifiers are not known until runtime, a lazy classifier such as K-nearest neighbors was used, achieving an overall accuracy of 89.7%. For the daily living activities recognition task, a bag of features approach was used. This method consists of modeling the entire activity as a distribution of physical activities called primitives. The average overall accuracy for five different daily living activities performed in an academic environment was 87.61%.
Master of Science in Intelligent Systems
Instituto Tecnológico y de Estudios Superiores de Monterrey
December 1, 2012
Repositorio Institucional del Tecnológico de Monterrey