Abstract

In road safety, real-time crash prediction may play a crucial role in preventing such traffic events. However, much of the research in this line generally uses data aggregated every five or ten minutes. This article proposes a new image-inspired data architecture capable of capturing the microscopic scene of vehicular behavior. In order to achieve this, an accident-prediction model is built for a section of the Autopista Central urban highway in Santiago, Chile, based on the concatenation of multiple-input Convolutional Neural Networks, using both the aggregated standard traffic data and the proposed architecture. Different oversampling methodologies are analyzed to balance the training data, finding that the Deep Convolutional Generative Adversarial Networks technique with random undersampling presents better results when generating synthetic instances that permit maximizing the predictive performance. Computational experiments suggest that this model outperforms other traditional prediction methodologies in terms of AUC and sensitivity values over a range of false positives with greater applicability in real life.