Information Extraction from Undersampled and Asynchronous Vehicle Sensor Data

• Main Author: Fox, Andrew S.
• Format: Others
• Published: Research Showcase @ CMU 2016
• Online Access: http://repository.cmu.edu/dissertations/749; http://repository.cmu.edu/cgi/viewcontent.cgi?article=1788&context=dissertations

As smart-vehicles, capable of advanced sensing and automated control procedures, become more prevalent in the Intelligent Transportation System (ITS), there is a need for next-generation road maps that contain all relevant environmental information that may assist drivers, passengers, and other stakeholders connected to the ITS. The proliferation of sensor-equipped consumer vehicles with dedicated communications systems has provided a valuable resource for continuous mobile data collection from which to extract such information. There are however, a number of challenges associated with such an information extraction process. Since vehicles and road environments are heterogeneous, there are many different possible sensor signals that could indicate an event. Sensor measurements may have significant error, which is often correlated within vehicles, particularly for measurements indexed by the Global Positioning System (GPS). Due to functional constraints of sensors, the measurements are asynchronously collected, and the signal-of-interest is often undersampled, requiring data to be aggregated from multiple vehicles to acquire a sufficient data set. In this thesis, we develop a vehicle-Cloud detection framework to extract environmental information from such aggregated, undersampled, and asynchronous vehicle data. We introduce the noisy multi-source, variable-rate (MSVR) sampling model with correlated errors in variables, and derive error models based on the vehicle and GPS sampling conditions. iv Road environmental information extraction algorithms are developed for the vehicle MSVR sampling conditions, specifically for the detection of continuous and binary types of road information. Within the overall vehicle-Cloud detection framework, algorithm adaptations are developed to detect events in multi-lane environments, filter data to reduce the required network bandwidth, account for temporally changing information, apply side information from other events, and use data-driven metrics to optimize the algorithm parameters. This framework is applied to detect road incline and bank angle information, and pothole locations on multi-lane roads. These algorithms are developed specifically for the MSVR sampling environment using only GPS and accelerometer data. Results are analyzed for sets of both simulated and real-world data, examining the tradeoffs between the number of aggregating vehicles and detection accuracy, in addition to the effects of the data filters and parameter optimizations developed in the overall detection framework.