Low-dimensional models for compression, estimation and prediction of large-scale traffic data

• Main Author: Mitrovic, Nikola
• Other Authors: Justin Dauwels
• Format: Theses and Dissertations
• Language: English
• Published: 2016
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/69423
• Institution: Nanyang Technological University

Intelligent Transportation Systems (ITS) often operate on large road networks and collect traffic data with high temporal resolution. The volume of the collected data severely limits the scalability of real-time traffic operations. We propose datadriven models that can help intelligent transportation systems in dealing with large amounts of traffic information in the context of online operations. We also propose non-centralized architectures for the implementation of the presented method on smartphones. Finally, we explore whether incorporating additional information (e.g., rainfall intensity) would lead to a better prediction and overall performance of the proposed system. Our main objective is to develop generic low-dimensional models that can enhance the scalability of traffic operations in practical city-scale networks, which are typically composed of different types of roads. We propose a column based (CX) matrix decomposition that leads to low-dimensional models where the components correspond to individual road segments (links) in the network. The resulting models can be easily interpreted and used for compression, compressed sensing and prediction. Hence, traffic states of large networks can be efficiently estimated by observing a small subset of road segments. To achieve this, we carefully select a small number of links from the original network. Then, we use historical information to learn the relationship functions between the selected subset of the links and the rest of the network. These functions allow us to estimate the traffic parameters (e.g., traffic speed) at any link in the network using information from a small subset of the links. Similarly, we perform traffic prediction for the whole network, by developing prediction models for only the representative subset of road segments. We model large networks by observing traffic conditions on a few roads and then extrapolating this information (through straightforward vector-matrix multiplication) across the network. This approach seems to be suitable for non-centralized architectures in which the workload is shared between central servers and user devices. To this end, we propose decentralized and hybrid architectures for system implementation, based on smartphones (as user devices) since they have remarkable performance nowadays. We study the computational performance of decentralized and hybrid architectures for applications of traffic speed estimation and prediction, and travel time prediction. We also analyze the performance of various development Android platforms and smartphone devices for different sizes of the test network. Weather conditions tend to have a significant impact on driving behavior and traffic performance. Assessing this impact might prove useful for the proposed lowdimensional models since they heavily rely on existing and future conditions at certain links in the network, which might be affected by inclement weather. However, quantifying this impact is not trivial. While light showers and a moderate amount of rainfall may not have any significant effect on road traffic, heavy showers, on the other hand, typically have a strong impact on driving conditions on roads. Furthermore, the resultant changes in traffic due to rain may vary during a day. We investigate the impact of rainfall on traffic conditions at different times of the day and for various rainfall intensities and road categories. In addition, we also investigate whether the incorporation of rainfall data would increase the performance of data-driven models for short-term traffic prediction. In summary, the proposed column-based (CX) method demonstrates a significant reduction in prediction time without significant degradation in prediction performance. Regarding the system implementation, the proposed decentralized and hybrid architectures provide flexibility for app developers and offer an opportunity to offload significant load from busy central servers. More importantly, the proposed decentralized architecture significantly reduces the overhead of the communication network and paves the way for new cooperative traffic applications and operations. Finally, the numerical results illustrate that, for certain links, incorporation of rainfall information would lead to better prediction performance of the traffic stream parameters.