Link Density Inference from Cellular Infrastructure

This work explores the problem of estimating road link densities from cellular tower signals by mobile subscribers in urban areas. We pose the estimation problem as a quadratic program, and present a robust framework that produces vehicle density estimates and is suitable for large-scale problems. We demonstrate that both simple and sophisticated models of cellular network connections can be handled robustly by the framework, without sacrificing efficiency or scalability. We present a numerical experiment on the I-15 corridor in San Diego based on a calibrated Aimsun microsimulation and a simulated cell network, demonstrating the framework can practically be implemented as part of an integrated corridor management system. The numerical results demonstrate that when the cell phone connection model is chosen appropriately, the estimates are consistent with those observed in a microsimulation. Yadlowsky, Thai, Wu, Pozdnukhov and Bayen 2 INTRODUCTION Vehicle density estimation is a critical component of future traffic management in urban areas. Vehicle density on freeways and arterial roads alike can be inferred from instrumentation added to the road network, but installing and maintaining such sensors is time consuming and costly, (Fontaine and Smith (1)). Heavily used freeways are usually instrumented, but most arterial roads are not. Thus, estimates of traffic conditions can only be found on a small fraction of the entire road network in an urban area. The framework presented in this article uses data from cellular network infrastructure to estimate current link densities on the road network. This not intended for traffic information systems, due to the granularity of the data, however better estimates of counts of cars on arterial roads on a 10-to-15 minute interval would be a useful input to traffic management systems that could provide richer information for decision making. For example, it could be used as data for improving demand estimation and management, and would allow traffic managers to adjust signal timing schedules to account for changes in current conditions, since most signal timing strategies are adjusted on a similar time resolution, (Lee and Williams (2)). Cell tower usage data has become an increasingly popular source of data for traffic demand estimation, as mobile phone network coverage is generally ubiquitous in urban areas, however one of the challenges of using cell phone tower data is the coarse granularity of the sensors, both spatially and temporally, (Cheng et al. (3)). For this reason, cellular infrastructure data is not as useful as GPS or Bluetooth for traffic information systems, (Herrera et al. (4) and Work et al. (5)). On the other hand, they are a pervasive source of data, where penetration rates in the population are exceptionally high compared to other data sources such as GPS or other wireless probes, as noted by Calabrese et al. (6). This makes them valuable for traffic management applications, where congestion and demand information are useful for making more data-driven management decisions. Studies of cell phone data have focused on numerous important areas of traffic modeling. Because of extensive coverage and appropriate level of accuracy and precision in cellular infrastructure data, applications to demand modeling have shown impressive results, allowing researchers to shift from census based models to more sophisticated models of data-driven origindestination inference, allowing higher temporal resolution, as presented by Toole et al. (7). In this article, we investigate the usage of this data to estimate current counts of vehicles on links in the road network, a more localized quantity of significant importance to traffic modeling. The map shown in Figure 1, created from the numerical work presented later in the article, shows how solving this problem is similar to projecting cell tower connection density onto the road network. In this article, we present a convex optimization framework for estimating link flows from cell tower usage data. Our model is based on two main ideas: connections to cell towers can be modeled by a probability distribution that gives the probability of being on each link given that one is connected to a given cell tower, and a link similarity model can encode the relationship in density between links in the road network. The framework presented poses the problem as a quadratic program that can be solved in O(n3) time, where n is the number of cell towers in the region, which is usually less than 1000. Numerical results establish consistency with microsimulation results, and discuss how this framework can be extended to further improve results. RELATED WORK Location data from mobile phones has been embraced by the urban planning community as a powerful and pervasive source of spatiotemporal data about urban communities, (Ratti et al. (8)). Yadlowsky, Thai, Wu, Pozdnukhov and Bayen 3 FIGURE 1 : Example of density projection from the data used for the present study. Regions (resp. links) colored in red have high cell tower connection (resp. vehicular) density. Best viewed in color. A time-lapse video is available on http://connected-corridors.berkeley.edu/