Advancing Android activity recognition service with Markov smoother

The rapid market shift to multi-functional mobile devices has created an opportunity to support activity recognition using the on-board sensors of these devices. Over the last decade, many activity recognition approaches have been proposed for various activities in different settings. Wearable sensors and augmented environments potentially have better accuracy, however performing activity recognition on user mobile devices has also attracted significant attention. This is because of less requirements on the environments and easier application deployment. Many solutions have been proposed by academia, but practical use is limited to testbed experiments. In 2013, Google released an activity recognition service on Android, putting this technology to the test. With its enormous market share, the impact is significant. In this paper, we present a systematic evaluation of this activity recognition service and share the lesson learnt. Through our experiments, we found scenarios in which the recognition accuracy was barely acceptable. To improve its accuracy, we developed ARshell in which we apply a Markov smoother to post-process the results generated by the recognition service. Our evaluation experiments show significant improvement in accuracy when compared to the original results. As a contribution to the community, we open-sourced ARshell on GitHub for application developers who are interested in this activity recognition service.