Avaliando modelos de predição de próximo sensor para trajetórias de veículos roubados

Abstract

The growing availability of trajectory data has driven the emergence of different techniques for predicting human mobility in recent years. Such data are collected either by applications installed on smartphones or by traffic monitoring systems in street networks. In this case, these systems are trajectories from external sensors (External Sensor Trajectories - EST) used in research and applications to the next location prediction. Record surveillance systems, among other property, restrictions on vehicle theft or theft. Knowing the dynamics of the movement of stolen vehicles in urban space is crucial information for government security agencies. However, the next location prediction (in this work, sensor) of a stolen vehicle is challenging due to the low regularity of transitions and the heterogeneity and scarcity of trajectory data. Therefore, this work offers a semantically enriched neural network model, analyzes the effectiveness of different machine learning models, investigates the best attributes for EST prediction models for stolen vehicles, and how different levels of spatial data representation can affect the prediction. We evaluate our model on a real dataset. The results demonstrated the effectiveness of machine learning models enriched with semantic data in predicting the next location in EST. The category of attributes related to Points of Interest contributed more than criminal data to most of the models tested, and that data representations with greater granularity favor the proposed solution.