towards an energy efficient transportation mode classification system

Continuous GPS data sampling on a GPS enabled mobile phone is expensive (.428 ~ .746 watts/sample) and drains battery quickly, which prevents users from using phones throughout a day without recharging. First, we want to investigate what sampling rate is needed, both to have satisfactory results for applications and to save battery life. High-level context such as bus routes and parking structure information and complex models such as hierarchical conditional random fields can be used to improve the performance of transportation classification techniques. For example, we are developing services that automatically infer that a user is riding a bus if his/her trace follows a bus route and pattern. However, although this approach shows significant improvements over existing techniques, it may be too heavy to run on mobile phones. Therefore, we are also investigating how simple models can be used without degrading the performance of systems. In a complementary investigation we observe that it is not necessary to collect fine-grained location data in some applications or situations. We previously demonstrated that an activity classifier using only coarse-grained location traces, GSM and WiFi beacons, successfully distinguishes meaningful differences in mobility states with 88% accuracy. We will continue to pursue this approach as an appropriate tradeoff between accuracy and longevity by evaluating its effectiveness using a larger set of data from multiple individuals and in a wider range of physical settings (public, residential, recreational, industrial, commercial); the latter is critical because the densities of GSM cells and WiFi access points vary significantly in different environments.