Stop detection for smartphone-based travel surveys using geo-spatial context and artificial neural networks

The problem of stop detection is at the base of many current and upcoming smartphone-based travel survey technologies and directly impacts the quality of many downstream operations. The inference of departure/arrival time, mode, and purpose of a trip, for example, rely on the stop/motion patterns represented by smartphone sensor-data. As users handle smartphones for various purposes and their preferences determine different device positions while traveling,
accelerometer and gyroscope, for instance, often present ambiguities that prevent accurate stop detection. To mitigate the impact of these ambiguities, we combine spatial time-series, i.e. GPS, with spatial context information retrieved from Open Street Map, which we represent as multidimension tensors. This project explores simple representations, such as dummy variables, and novel multidimensional representations, which are bench-marked through the classification performance of specialized artificial neural network (ANN), as well as other machine learning (ML) baselines. Our main contribution stems from this novel multidimensional representation of time-series fusion with spatial context, combined with the corresponding specialized ANN classifier. The results show a stop detection score improvement on the baselines between 3% and 6.5%.