Date of Degree


Document Type


Degree Name



Computer Science


Hanghang Tong

Committee Members

Ted Brown

Qing He

Ping Ji

Zhigang Zhu

Subject Categories

Artificial Intelligence and Robotics | Other Computer Sciences


Mobile Mining, Machine Learning, Travel Mode Detection, Personalized Learning, Smartphone Sensors


Personal trips in a modern urban society typically involve multiple travel modes. Recognizing a traveller's transportation mode is not only critical to personal context-awareness in related applications, but also essential to urban traffic operations, transportation planning, and facility design. While the state of the art in travel mode recognition mainly relies on large-scale infrastructure-based fixed sensors or on individuals' GPS devices, the emergence of the smartphone provides a promising alternative with its ever-growing computing, networking, and sensing powers. In this thesis, we propose new algorithms for travel mode identification using smartphone sensors. The prototype system is built upon the latest Android and iOS platforms with multimodality sensors. It takes smartphone sensor data as the input, and aims to identify six travel modes: walking, jogging, bicycling, driving a car, riding a bus, taking a subway. The methods and algorithms presented in our work are guided by two key design principles. First, careful consideration of smartphones' limited computing resources and batteries should be taken. Second, careful balancing of the following dimensions (i) user-adaptability, (ii) energy efficiency, and (iii) computation speed. There are three key challenges in travel mode identification with smartphone sensors, stemming from the three steps in a typical mobile mining procedure. They are (C1) data capturing and preprocessing, (C2) feature engineering, and (C3) model training and adaptation. This thesis is our response to the challenges above. To address the first challenge (C1), in Chapter 4 we develop a smartphone app that collects a multitude of smartphone sensor measurement data, and showcase a comprehensive set of de-noising techniques. To tackle challenge (C2), in Chapter 5 we design feature extraction methods that carefully balance prediction accuracy, computation time, and battery consumption. And to answer challenge (C3), in Chapters 6,7 and 8 we design different learning models to accommodate different situations in model training. A hierarchical model with dynamic sensor selection is designed to address the energy consumption issue. We propose a personalized model that adapts to each traveller's specific travel behavior using limited labeled data. We also propose an online model for the purpose of addressing the model updating problem with large scaled data. In addressing the challenges and proposing solutions, this thesis provides an comprehensive study and gives a systematic solution for travel mode detection with smartphone sensors.