Mobility modeling of mobile users.

Abstract

Accurate modeling of the movement of mobile users is important for the design and operation of mobile systems, because (1) user mobility model can reproduce movement in simulating a mobile system for performance evaluation, (2) accurate knowledge on user movement can help researchers develop and improve location-/movement-aware applications, and (3) it can help us efficiently manage and optimize existing mobile networks in service. However, tracking real people for modeling is very hard due to privacy, cost, and technical concerns. While many random models have been constructed based on simple heuristics, they may or may not represent real movement of mobile: users. With this motivation, this thesis focuses on the realistic modeling of user mobility. We first take on a very widely used synthetic random mobility model, and show that the way it has been commonly used is fundamentally flawed. The problem is attributed to the non-stationarity of the model, which leads to misleading and unreliable results. We analyze the origin of the problem and present a general framework within which stationary versions of these mobility models may be constructed in a systematic way. We then turn our attention to trace data of real mobile users. Using the coarse-grained association logs of mobile users in a campus-wide wireless network at Dartmouth College, we propose a methodology to generate a mobility model that is realistic, site-dependent, tailored to a specific time period, and easily updatable. The resulting model is evaluated against real measurements of pedestrians at intersections on campus. Continuing to work with real movement data, we turn to more fine-grained GPS trace data collected from moving vehicles to identify traffic status. Using spatial/temporal traffic features from GPS data and our unique threshold-based clustering algorithm, we show how identification of current traffic condition and anomaly detection can be done. The evaluation with two different trace sets shows that higher than 90% of accuracy can be achieved after short system training. Moreover, we show that traffic patterns on roads are highly consistent over time that we can use more data of a longer history for higher accuracy.