Mining smartphone generated data for user action recognition – Preliminary assessment

Smartphones became everyday “companions” of humans. Almost everyone has a smartphone in their pocket, or bag, and use it on daily basis. Modern smartphones are “loaded” with sensors, providing streams of, potentially useful, data. Simultaneously, staying fit, exercising, running, swimming, etc. became fashionable. In this “climate”, employers can try to incentivise their workers, for instance, to use bicycles to come to work. Here, one of interesting questions becomes: are workers actually using bicycles, as declared, or do they try to subvert the system and win prizes, while, for instance, using public transport. One of the ways to check this could be to use data from smartphone sensors to determine the mode of transportation that has been used.

This paper presents preliminary results of an attempt at using raw sensor data and deep learning techniques for transportation mode detection, in real-time, directly on smartphone. The work tries to balance sensor power consumption and computational requirements with prediction correctness and response time. In this context, results of application of recurrent neural networks, as well as more traditional approaches, to a set of actual mobility data, are presented. Furthermore, approaches that leverage domain knowledge, in order to make classifiers more reliable and requiring less processing power (and less energy), are considered.