DISTRIBUTED COMPUTING AND UNSUPERVISED DEEP LEARNING FOR ANALYZING HUMAN TRAVEL BEHAVIORS USING BIG TRAJECTORY DATA

Abstract

Human travel behaviors, which refer to the movement of either individuals or groups across space and time, has been a research focus for the understanding of not only the behaviors themselves, but also the interaction between humans and their environment. In recent years, with the development of information and communication technologies, passively collected trajectory data (i.e., sequences of sampled geographic coordinates with timestamps collected by GPS sensors in devices such as mobile phones and vehicles) has become more available. This new data type has provided unprecedented opportunities for research in human travel behaviors offering advantages of broad spatial scale, large data volume, and fine granularity compared with traditional actively collected data, such as census data and survey data. This dissertation focused on exploring how advanced research methods, including distributed computing and unsupervised deep learning, can be extended and applied to examine human travel behaviors using passively collected trajectory data. In this dissertation, three studies are described that examine three different topic that contribute to our understanding of vehicular travel behaviors. The first study involved developing a distributed research framework based on Apache Spark and Sedona to estimate traffic speed and speeding across a state-wide road network and multiple road types using passively collected mobile device data. The study analyzes spatio-temporal patterns in traffic speed and speeding behaviors in the state of California and examines differences in patterns from different road types such as freeways and residential roads. The second study developed a research framework that combines a rule-based distributed parking extracting module and an LSTM-autoencoder model to extract parking trajectories and automatically classify them into different categories (e.g., direct parking and cruising for parking) using in-vehicle trajectory data. Using datasets collected in the Washington, DC metropolitan area, this study further investigates the driving patterns and spatial distribution of cruising trips, identifying areas where parking demand may exceed supply. The third study in the dissertation proposed a research framework that is comprised of a distributed contour plot-building module (designed to capture speed changes compared with free-flow and historical averages), a W-net-based semantic segmentation model to segment the non-recurrent congestion (NRC) impact areas in these plots, and a postprocessing module to ensure that the identified propagation patterns follow the law of shockwaves using passively collected in-vehicle trajectory data. Based on detected NRC impact areas, propagation of NRC over actual road networks in the Washington, DC metropolitan area was examined to identify how these types of congestion move through these road networks. The contributions of this dissertation include providing cutting-edge distributed computing and deep learning methods to make full use of passively collected trajectory data for human travel behavior analyses and provide feasible means to overcome the challenges brought by the large data volumes, broad spatio-temporal ranges, and lack of ground truth labels. The integration of passively collected trajectory data and new research methods have provided new insights into human travel behaviors over large spatial scales and fine granularities, and lay the foundation for future research on human mobility.