Lightweight Semantic Location and Activity Recognition on Android Smartphones with TensorFlow

Context-awareness is the parent concept that guides the applications we will discuss here. This concept is tightly related with ubiquitous computing, a term that encloses all those technologies that have pervasively made their way in countless aspects of our everyday life. What we want from this work is to find a way to approach the problem of Semantic Location detection and Activity Recognition, addressing two of the main factors in pollutants exposure: involved activity and microenvironment. We focus this work on exploiting the possibilities that mobile device can offer today, motivated by their pervasive presence in our daily life, their always increasing technological capabilities and their diffusion: more than seventy-five billion devices will be connected to the internet by the year 2025—this is ten times the Earth population today. We use onboard sensors on common smartphones to collect motion data for activity recognition, and other sensors and similar information to determine indoor or outdoor positioning. We exploit the following smartphone sensors and components. This approach is supported and inspired by plenty of research literature and tailored to our specific needs. We work with Android devices and aim to build a prototype for a mobile application that can collect this data from the device and run a machine learning based model to run inference locally. We introduce a less known version of neural networks called Long Short-Term Memories that are uniquely intended to work with temporal data, and that today is employed by most of the major technology companies out there for their top tier products involving machine learning.