Modelos baseados em PPM para previsão de trajetórias utilizando informações contextuais.

Abstract

Thanks to the widely diffusion of smartphones with GPS devices natively embedded, the task of tracking object locations, such as people or vehicles, is more feasible nowadays, fostering new research opportunities. Among these new opportunities, this work addresses the challenge of route and destination prediction. Knowing in advance the destination where a user might reach as soon as he or she starts to move can be useful in various situations. For instance, to suggest to users less jammed or safer routes, as well to warn about points of interest located along their route. There are commercial systems capable of predicting destination and routes, however, these systems usually require frequent user interaction. Nonetheless, such a requirement could make the application unusable for daily routines. Moreover, most existing works do not consider an important contextual information: the information about the places that the users visit, i.e., the role that the places play to the user (for instance, if the place is home or work). In addition, most predictors described in the literature are not able to predict places that users have never visited. This thesis proposes a family of algorithms based on Prediction by Partial Matching (PPM). Furthermore, this work proposes a mechanism for identifying whether a route is being performed for the first time, resulting in the feasibility for predicting a never visited place. This research also provides a comparison between our proposed predictors, and the predictors based on Markov Models and Hidden Markov Models (HMM), which have been used in related works. It is important to mention that both Markov and HMM predictors that we implemented are able to predict just the destination, instead the remaining route. For the statistical assessment of the predictors, the metrics Precision, Recall and F1 Score are used, together with the process of 10-fold cross- validation. The database contains about 1,500 routes extracted from 21 users, gathered for three months. The predictors based on PPM performed similarly (or better) than others reported in the literature.