Kernel-based estimation of individual location densities from smartphone data

Localizing people across space and over time is a relevant and challenging problem in many modern applications. Smartphone ubiquity gives the opportunity to collect useful individual data as never before. In this work, the focus is on location data collected by smartphone applications. We propose a kernel-based density estimation approach that exploits cyclical spatio-temporal patterns of people to estimate the individual location density at any time, uncertainty included. Model parameters are estimated by maximum likelihood cross-validation. Unlike classic tracking methods designed for high spatio-temporal resolution data, the approach is suitable when location data are sparse in time and are affected by non-negligible errors. The approach is applied to location data collected by the Earthquake Network citizen science project which carries out a worldwide earthquake early warning system based on smartphones. The approach is parsimonious and is suitable to model location data gathered by any location-aware smartphone application.