| Summary: | Understanding urban mobility is becoming one of the most critical tasks as more of the world’s cities become crowded and congested. Given the increasing volume and variety of urban mobility, traditional survey based methods are insufficient to sense the whole picture, based on which few need-satisfying solutions can be built. Fortunately, over the past few years, we also experienced a significant increase in peoples ability to collect data from various sensors, smartphones, or other devices, in different formats. During my PhD program, I focus on designing holistic systems for sensing and analyzing mobility for urban transportation. I aim at mining insightful mobility information from not only dedicated sensors but also from smartphones, wearable devices and city infrastructures like traffic cameras. In this thesis, I present three studies for mobility sensing and analysis of different subjects and scale in urban context.
The first study considers the mobility of individual smartphone user. I observed that although various algorithms have been proposed to infer a user’s location from data sensed by the smartphone, none of existing localization algorithms can works accurately in all environment, which results in incomplete or noisy mobility. To address this problem, I propose UniLoc, a unified framework that exploits the scheme diversity to gain extra performance improvement from state-of-the-art localization solutions. In UniLoc, multiple localization schemes are executed in parallel and independently. UniLoc predicts the localization accuracy of each scheme online according to the real-time context at per-location granularity. In a simple version, UniLoc chooses the best scheme as its final result at every location. Beyond that, UniLoc combines the results of all available schemes based on a locally-weighted Bayesian model averaging algorithm. The combined result is better than any individual scheme. Extensive experiments demonstrate that such an easy aggregation incurs little overhead in energy consumption or training a new scheme, but gains substantially from the localization scheme diversity.
The second study considers mobility of thousands of students. I found that collective mobility information can be crowdsensed from wearable devices and better urban services can thus be planed via proper analysis. By processing a large dataset composed of daily trajectories of thousands of students in Singapore, I find that, instead of simply picking up students from their homes, an optimal school shuttle planning system needs to learn the real transportation usage and plan across all potential pickup locations for every student to generate need-satisfying routes. It is challenging, however, to perform route planning over a large number of students each having multiple potential pickup locations. We develop a graph-based data structure that embeds potential pickup locations of all students with the awareness of real-world constraints and existing public transits. Based on the graph structure, we prove that the optimal last-mile school shuttle planning problem is NP-hard and thereafter design a Tabu-based expansion algorithm to solve the problem, which strikes at a proper balance between the savings of students’ commute time and the total cost of operating the shuttle buses. Extensive experiments with large-scale real-world crowdsensed trajectory data demonstrate that our last-mile school shuttles can save the traveling time for most students by over 20% and the savings can be up to 65% for 10% of the students.
The third study considers the mobility of all general vehicles in a city. Vehicle trajectories provide fundamental understanding of the urban mobility and benefit a wide range of urban applications. However, state-of-the-art solutions for vehicle sensing (i.e., tracking with in-vehicle devices, roadway monitoring with flow sensors) can only acquire either partial or anonymized observations, thus fails to reconstruct all individual vehicle trajectories. To address those drawbacks, We propose VeTrac, a comprehensive system that uses widely deployed traffic cameras as a sensing network and reconstructs the complete knowledge of all general vehicle in the city. Despite recent advances in identifying vehicles based on their plates or appearance, we observe substantive vehicle identity uncertainties from existing methods. Accordingly, we design VeTrac that exploits the multi-dimensional similarity and embeds that in a graph convolution process in order to reduce the uncertainties of vehicle identification. VeTrac also employs a self-training process that incorporates the complex mobility dependency among the urban road networks to improve accuracy. Extensive experiments demonstrate that VeTrac reconstructs trajectories with an overall 89% accuracy and outperforms existing algorithms.
|
|---|