Construcción de un dataset de eventos de conducción utilizando modelos de generación de datos sintéticos mediante Generative Adversarial Networks (GAN).

Abstract

The development of solutions to reduce traffic accidents in vehicles requires a lot of experimentation and, therefore, a large amount of data. However, due to data confidentiality issues, few free datasets are available in this area, so it is necessary to build datasets that contain several attributes and can be of general use in this field. Given this reality, this project uses the TimeGAN framework for synthetic data generation. TimeGAN is configured in this thesis work with four recurrent neural networks, which obtain a spatial and temporal distribution that seeks to approximate the real data. The synthetic data generated can replace or complement the real data in different scenarios, with the advantage that the synthetic data can be shared without the risk of exposing confidential information of persons or entities associated with these data. The synthetic data will serve to complete the information of an initial dataset built with driving data obtained from heterogeneous sources. In order to evaluate the results of synthetic data generation, qualitative and quantitative methods will be used to establish the usefulness of the data. The synthetic data dataset constructed will then be used in processes of classification of driving events that present a higher risk of causing traffic accidents.