Probabilistic models for mobile phone trajectory estimation

This dissertation is concerned with the problem of determining the track or trajectory of a mobile device — for example, a sequence of road segments on an outdoor map, or a sequence of rooms visited inside a building — in an energy-efficient and accurate manner. GPS, the dominant positioning technology today, has two major limitations. First, it consumes significant power on mobile phones, making it impractical for continuous monitoring. Second, it does not work indoors. This dissertation develops two ways to address these limitations: (a) subsampling GPS to save energy, and (b) using alternatives to GPS such as WiFi localization, cellular localization, and inertial sensing (with the accelerometer and gyroscope) that consume less energy and work indoors. The key challenge is to match a sequence of infrequent (from sub-sampling) and inaccurate (fromWiFi, cellular or inertial sensing) position samples to an accurate output trajectory. This dissertation presents three systems, all using probabilistic models, to accomplish this matching. The first, VTrack, uses Hidden Markov Models to match noisy or sparsely sampled geographic (lat, lon) coordinates to a sequence of road segments on a map. We evaluate VTrack on 800 drive hours of GPS and WiFi localization data collected from 25 taxicabs in Boston. We find that VTrack tolerates significant noise and outages in location estimates, and saves energy, while providing accurate enough trajectories for applications like travel-time aware route planning. CTrack improves on VTrack with a Markov Model that uses “soft” information in the form of raw WiFi or cellular signal strengths, rather than geographic coordinates. It also uses movement and turn “hints” from the accelerometer and compass to improve accuracy. We implement CTrack on Android phones, and evaluate it on cellular signal data from over 126 (1,074 miles) hours of driving data. CTrack can retrieve over 75% of a user’s drive accurately on average, even from highly inaccurate (175 metres raw position error) GSM data. iTrack uses a particle filter to combine inertial sensing data from the accelerometer and gyroscope with WiFi signals and accurately track a mobile phone indoors. iTrack has been implemented on the iPhone, and can track a user to within less than a metre when walking with the phone in the hand or pants pocket, over 5× more accurately than existing WiFi localization approaches. iTrack also requires very little manual effort for training, unlike existing localization systems that require a user to visit hundreds or thousands of locations in a building and mark them on a map. Thesis Supervisor: Hari Balakrishnan Title: Professor of Computer Science and Engineering Thesis Supervisor: Samuel R. Madden Title: Associate Professor of Computer Science and Engineering