Abstract:
We look at the problem of using accelerometer in smartphones to detect whether
the user is at a metro train station or in a metro train. Currently, we have
solutions to detect simple activities, such as sitting or walking. Our work for
this thesis investigates the more complex problem of discerning \in-train" from
\in-metro-station" activities which internally are composed of several simple activities.
We perform the task of distinguishing the \in-train" from \in-metrostation"
patterns using classic classification techniques with two different data
representations namely, statistical features and ECDF-based features. Another
major contribution through this thesis is to solve the challenge of the considerable
class imbalance with majority of samples belonging to the \in-train" patterns by
improvising existing classification algorithms to counter the effect of class imbalance.
Our fndings are useful for any other problem of using sensor data to classify
activities. We evaluated our solution using about 23 hours of data collected using
six different models of smartphones from over seven different metro/subway
stations situated in New Delhi, India. Our detection accuracy is over 98%.
