Detecting activities at metro stations using smartphone sensors

Our paper aims to build a classification-model which delineates the typical motion-related activities performed at a metro station using smart phone sensors. We focus on typical movements, such as climbing the stairs or moving in the lift, waiting at security, waiting at the turnstile to check out and, moving on platform while waiting for a train. Such a classifier estimates crowd levels in a metro-station (and metro trains in an indirect sense), thereby adding towards the vision of efficient metro travel. However, building such a classification-model is challenging due to non-trivial decision boundaries among the classes of interest. Our experiments revealed that the best accuracy that a traditional multi-class classifier could obtain was 0.58, for a four-class classifier. To this end, we proposed a hierarchical approach of classification which divides the multi-class problem at hand into a set of manageable two-class classification problems. These two-class classifiers are then put together, in a hierarchy, to give an end-to-end solution which takes sensors values from the phone and predicts the class of motion being observed. Our model obtained an accuracy of around 0.75, a significantly higher value, on the real-data collected at Delhi Metro stations. The same classifiers can potentially be applied to detect crowd levels at train stations and bus depots, which will make transportation efficient in smart cities.