Automating and Optimizing a Transportation Mode Classification Model for use on Smartphone Data
Loading...
Date
2015-04-01
Authors
Nour, Akram
Journal Title
Journal ISSN
Volume Title
Publisher
University of Waterloo
Abstract
As transportation engineering and planning evolve from “data poor” to “data rich” practices, methods to automate the collection and translation of data to information become increasingly important. Advances in wireless communications and technologies provide the opportunity to collect detailed data on travel trajectories using smartphones equipped with GPS and accelerometers. These types of smartphones are ubiquitous and, as such, present an opportunity to conveniently collect spatial and temporal data at regular time intervals. This can be useful to utilize as a method to document trip attributes of interest, namely origin, destination, departure time, route choice, trip purpose, and mode choice. Though some of these attributes can be relatively easily extracted from the smartphone data, inferring transportation mode(s) used by the trip maker remains a challenging problem.
This research presents a data-driven classification model to infer the transportation mode(s) used by trip makers on the basis of data collected with GPS equipped smart phones. Rather than making a priori assumptions, we instead employ an optimization method to objectively produce the following classifier components and methods: a ranked feature vector based on the power of differentiation between different modes; the classification technique between the range of candidate classifiers; the number of ranked attributes to include in the feature vector; data formatting; and optimal model parameters. The model is trained and tested using labelled trip data. The calibrated model is evaluated by testing its ability to classify travel mode correctly for GPS data at a different level of disaggregation than the one used in the model training step. The model provides an accuracy of approximately 86% at the disaggregated level (e.g., Walk, Bike, Transit, and Private Automobile) and approximately 94% at aggregated level (e.g., Non-Motorized and Motorized.)
The results obtained from the optimized model are supplemented with a GIS based model to improve the identification of transit trips. The method employed integrates GIS data such as the locations of transit stops and signalized intersections with observed travel patterns from the GPS embedded smartphone data. The combination of these two data sources generates new classification features that, when applied to the collected data, demonstrate that this technique vastly improves the accuracy of the classification model for identifying transit mode usage.
Description
Keywords
Travel Behaviour, Classification model, Machine Learning, Transportation Mode Identification, GPS, GIS, Spatial Statistics, Transportation Planning, ITS, Optimization, Intelligent Transportation Systems, Smartphone Data, Data Analysis, Transportation Mode, Modeling, System Development