We analyze the ability of mobile phone-generated accelerometer data to detect high-level (i.e., at the semantic level) indoor lifestyle activities, such as cooking at home and working at the workplace, in practical settings. We design a 2-Tier activity extraction framework (called SAMMPLE) for our purpose. Using this, we evaluate discriminatory power of activity structures along the dimension of statistical features and after a transformation to a sequence of individual locomotive micro-activities (e.g. sitting or standing). Our findings from 152 days of real-life behavioral traces reveal that locomotive signatures achieve an average accuracy of 77.14%, an improvement of 16.37% over directly using statistical features.