Montagem de infraestrutura e predição de trajetória em redes veiculares

Abstract

Vehicular Ad-Hoc Networks (Vanets) are networks composed by vehicles within sensing and communication capabilities and that exchange messages among themselves or among access points deployed in the region. Data collected by Vanets offers, among others services, information about conditions of roads, traffic and climate; the behavior of vehicles and drivers. Besides vehicles, access points are the main agents of information dissemination, that help to overcome some communication limitations of Vanets. Thus, a study to deploy a Vanet infrastructure is very important, so the exchange of information quality is facilitate and improved. Localization and tracking techniques allow knowing the current position of a vehicle, hence the applications can be more interesting whereas they can be directed and adapted to the environment and/or user involved. Thus, some inference about the driver behavior can be done, for instance, next positions, trajectories and lane changes. Localization data of vehicles can be used to deploy the infrastructure of a Vanet, because if data belongs to vehicles that are moving in a common region it is possible identify the global traffic behavior and so distribute access points to improve the communication quality of the network. Moreover, this data is primordial to make trajectory predictions of vehicles, because from the history is possible to identify driver patterns and apply some techniques that is able to make some future behavior from this information. Concerning to the problem of installing infrastructure, this works proposes a genetic algorithm to distribute access points in a region to reach the best vehicle coverage. Regarding to the customization of the data dissemination, is presented a modeling, applicable to classic algorithms from machine learning, to predict trajectories of vehicles. The genetic method, to deploy the infrastructure, was applied in four scenarios with real topology of Switzerland roads, considering a realistic mobile vehicular model during one hour and a half. Results show that the genetic algorithm, with a population initialization method that explores some solutions generated by a greedy approach and with genetic operators developed with problem inherent information, presents solutions up to 20.12 pp better than the greedy solution. Other results, varying the number of access points available and the minimum time of information receipt, show that the genetic algorithm always beats the greedy solution and the random greedy. About the prediction problem, a real data set collected from the Borlange city, in Sweden, is used. This data set has 24 users, with different characteristics. This set was adapted to apply to some classification algorithms employing the sliding window concept.A quantitative study of the data set was presented and an analysis of its behavior with four learning algorithms coded on the scikit-learn framework: (i) k-Nearest Neighbors, (ii) Naive-Bayes, (iii) SVM, and (iv) Decision Tree. Each vehicle route is modeled as a graph and the objective is, given a sequence of edges, predict the next edge. Results show that the decision tree classifies successfully 0.85 of the instances.