Semantic Activity Classification Using Locomotive Signatures from Mobile Phones

We explore the use of mobile phone-generated sensor feeds to determine the high-level (i.e., at the semantic level), indoor, lifestyle activities of individuals, such as cooking & dining at home and working & having lunch at the work- place. We propose and evaluate a 2-Tier activity extraction framework (called SAMMPLE1) where features of the low-level accelerometer data are first used to identify individual locomotive micro-activities (e.g., sitting or standing), and the micro-activity sequence is subsequently used to identify the discriminatory characteristics of individual semantic activities. Using 152 days of real-life behavioral traces from users, our approach achieves an average accuracy of 77.14%, an improvement of 16.37% from the traditional 1-Tier approach, which directly uses statistical features of the accelerometer stream, towards such activity classification tasks.