Framework for Human Activity Recognition on Smartphones and Smartwatches.

As activity recognition becomes an integral part of many mobile applications, its requirement for lightweight and accurate techniques leads to development of new tools and algorithms. This paper has three main contributions: (1) to design an architecture for automatic data collection, thus reducing the time and cost and making the process of developing new activity recognition techniques convenient for software developers as well as for the end users; (2) to develop new algorithm for activity recognition based on Long Short Term Memory networks, which is able to learn features from raw accelerometer data, completely bypassing the process of generating hand-crafted features; and (3) to investigate which combinations of smartphone and smartwatch sensors gives the best results for the activity recognition problem, i.e. to analyze if the accuracy benefits of those combinations are greater than the additional costs for combining those sensors.