The next transportation revolution relies on the availability and sophisticated processing and interpretation of extensive transportation data. The first bottleneck, data collection, is no longer a challenge. If anything, we now have a massive amount of transportation data, including data on driver behavior, traffic, weather, accidents, vehicle messages, traffic signals and road characteristics. Our next challenge is how to make such data accessible to researchers and to transportation regulation agencies, and how we make sense of the data, to gain insight for future transportation research, the development of connected and automated vehicle systems, and regulations and guidelines in response to new transportation systems.
This project, led by Carol Flannagan of the U-M Transportation Research Institute, will create a data ecosystem in a high-performance computing environment to understand driver behavior, and demonstrate its value through its application to a practical research issue and the development of a specific set of analytical methods. First, the research team will build the data ecosystem on a parallel, distributed computing platform that is optimized to support a variety of Big Data analytical methods and data integration across a variety of data sources, empowering researchers and transportation agencies. The team will then demonstrate the value of the data ecosystem by developing a database of rare or unusual traffic events and a rich and natural simulation environment for such events. It is often ineffective for technology developers and regulators to test designs and regulations in response to certain rare transportation events through road tests, because these rare events are, well, rare. An enriched simulation setup will allow much more effective testing. The research team will also develop methods to integrate Information collected by different sensors, with different levels of details and quality, and across different time scales. Success of such method development not only proves the value of a particular data ecosystem, but also has broad implications in the advancement of data-intensive research in many fields because the integration of signals of different origins, qualities, levels of details and time scales is a challenge that faces researchers in many fields such as patient data analysis and social media data analysis.
The research team include researchers from the School of Public Health, College of Engineering, U-M Dearborn, College of Literature, Science and the Arts, UMTRI, and the Institute for Social Research, a perfect combination of expertise to carry out the proposed multi-disciplinary research.
- Carol Flannagan’s slides (PDF) from the Midwest Big Data Hub Transportation Symposium held on the U-M campus in June 2017
Carol A. Flannagan, Principal Investigator, U-M Transportation Research Institute
Michael R. Elliott, Biostatistics
Robert Hampshire, Ford School of Public Policy
H.V. Jagadish, Computer Science and Engineering
Judy Jin, Industrial and Operations Engineering
Jason Mars, Computer Science and Engineering
Aditi Misra, UM Transportation Research Institute
Yi Lu Murphey, Electrical and Computer Engineering, UM-Dearborn
Kerby Shedden, Department of Statistics
Lingjia Tang, Computer Science and Engineering
Kristine Witkowski, Institute for Social Research
- The team has secured further funding from Toyota Research Institute, Ford Alliance and ZF.
- The team has piloted a video and location analytic system for driver behavior, scene recognition and object detection. The system is deep-neural-network based, fast and scalable, and accepts intuitive queries.
- The team has developed a novel streaming data methodology for analyzing driver behavior using high frequency sensor data.
- The team has developed methods to improve the representativeness of non-probability samples by comparison with probability samples.
- The team has developed a number of Machine Learning algorithms to detect distracted driving and driver fatigue based on vehicle kinematics and video data.
The team has set up a baseline computing system for running computer vision algorithms on driving data and sensor data. We have run an object-detection algorithm on a subset of the data and are working on speed processing and optimizing the data videos. In addition, the team is refining methods to produce nationally representative results from the analysis of Safety Pilot data, and identify feature-specific (such as safety) epochs in the data for further analysis.
The next steps include further experiments to speed up video analyses, experiments in improved GPS privacy algorithms, and additional applications. The data and algorithms will be available to the U-M community next year. (July 2017)