The objective of this project was to explore and harness the emerging automotive infoverse in order to achieve a variety of performance, efficiency, reliability, comfort, and safety (PERCS) goals through high-confidence vehicle tuning and driver feedback decisions. The automotive infoverse captures extrinsic information in addition to conventional short-term intrinsic vehicle sensor data. It’s key sources of information are historic triplines, socialized versions of these that are shared amongst vehicle users, and online databases. Specifically, the goals of the project are the following. First, developing software infrastructure that permits the rapid development of apps that use the infoverse to achieve one or more PERCS goals. Second, exploring the design of infoverse apps such as advisory apps that can provide feedback to the driver in order to improve vehicular safety, or auto-tuning apps that can safely set car parameters to achieve a PERCS goal. Allowing vehicles and such apps to share vehicle data with others and to use extrinsic information in their decisions results in novel information processing, assurance, and privacy challenges. A third goal or this project is therefore to develop methods, algorithms and models to address these challenges and enable harnessing the automotive infoverse.
To address the first goal, the project developed CarLog infrastructure and set up a testbed in an effort led by our project partners at the University of Southern California. CarLog is a programming system that simplifies the development of event-driven automotive apps. We observe that a declarative logic-based language like Datalog is an excellent choice for programming in this setting. We further developed a testbed for crowd-sourced automotive sensing: a software framework that would allow multiple users to automatically collect and upload sensor information to the cloud, with minimal user intervention. The testbed consists of, in essence, a mobile application (app) and a set of web services. The app, for the purposes of our experiments serves as a means to probe for and transmit car data from the car’s onboard sensor systems to a centralized data repository. The testbed was deployed on 10 mail vans at USC and collected sensor data continuously over much of the project duration. This resulted in a dataset of over 500 million sensor readings from this fleet.
The project then built on this infrastructure to explore a variety of sensing, positioning, robustness, and privacy challenges. It explored the opportunities of fusing in-vehicle sensor data with those from occupants’ smartphones. Our findings suggest that neither car-sensing nor phone-sensing is strictly better than the other, but that crowd-sourcing is essential for both. It demonstrated that fusing inertial tripline data maps and Global Positioning System measurements can yield improvements in vehicle positioning accuracy. It also showed that the robustness of GPS vehicle positioning can be improved through crowd-sourced interference detection. The projects then explored sharing of data from richer sensors such as camera or Lidar and showed applications in road traffic flow estimation as well as how it can realize a concept we termed augmented vehicular reality. By exchanging rich point cloud data among vehicles and merging received data with their locally sensed data, vehicles can monitor obstacles or traffic participants that their own sensors cannot detect, for example due to occlusion.
We finally demonstrated techniques to collect critical data for improving the robustness of driver assistance systems and emerging automated driving systems. A key challenge here is understanding not just common situations of highway and city traffic, but also the long tail of diverse and unusual traffic events that such systems need to handle. Capturing such events usually requires driving significant testing miles. To address this challenge, we study an automatic unusual driving event identification system, which can detect and capture unusual situations through in-vehicle algorithms from a much larger fleet of human-driven vehicles.
The project has enabled six graduate students (3 from USC, 3 from Rutgers), two faculty members (one each at USC and Rutgers), and researchers at General Motors to collaborate with each other and understand the information infrastructure in modern automobiles as well as understand other aspects of transportation related cyber-physical systems. With this project, our students will be trained to advance the cause of transportation safety and efficiency, and help develop improved transportation infrastructures using advanced cyberphysical technologies. It helped three graduate students, two of them from underrepresented groups, complete a thesis. Moreover, the team founded and organized the ACM International Workshop on Smart, Autonomous, and Connected Vehicular Systems and Services (CarSys) held with the ACM MobiCom conference to facilitate the exchange of results and provide a forum for researchers in this area. We hope that these outcomes can facilitate safer and more efficient automotive transportation across a variety of stages of vehicle automation.