A Data-Driven Cyber-Physical System for Urban Mobility Modeling and Their Applications

Abstract

Nowadays, we are in a rapid process of urbanization, which leads to severe mobility challenges, e.g., traffic congestion and gas consumption. To address these challenges, it is essential to model human mobility, and improve urban mobility efficiency with novel applications based on mobility models. Existing human mobility models and resultant applications are mostly driven by data from isolated urban systems, e.g., cellphone networks or transportation systems, which leads to a bias against urban residents not involved and thus inefficiency of resultant applications. In this dissertation, we propose a cyber-physical system called mobileCPS to model the human mobility at fine spatiotemporal granularity and then design novel mobility-driven applications. Specifically, we design a three layer architecture for mobileCPS: (i) a real-time data feed layer where we collect multi-source urban data related mobility from extremely-large urban infrastructures, e.g., cellphone networks and transportation systems, which is one of the largest urban data consolidations for academic research; (ii) a mobility abstraction layer where we design a human mobility model driven by multi-source data we collected with a multi-view learning technique, which is the first work that models human mobility with multi-source data; (iii) an application design layer where we present two mobility driven applications, i.e., a real-time carpooling service called coRide and last-mile transit service called Feeder, to improve urban mobility efficiency. coRide is the first systemic carpooling service with real-world implementation and a dynamic fare model, and Feeder is the first last-mile transit service driven by multi-source urban data. The key intellectual contributions of this work include (i) a human mobility modeling technique iteratively driven by heterogenous multi-source urban data; (ii) a set of optimal, approximation and online algorithms for a mobility-driven carpooling problem; (iii) a data-driven inference technique for last-mile passenger demand. We implement and evaluate mobileCPS based on extremely large datasets in the Chinese city Shenzhen with cellphone and transportation systems including taxis, buses, and subways, capturing more than 27 thousand vehicles and 10 million urban residents. The results show that mobileCPS (i) increases mobility model accuracy by 51%, (ii) reduces mileage by 33% with its carpooling, and (iii) reduces the last mile distance by 68%.