Predicción de niveles de accidentabilidad en accidentes de tránsito usando redes neuronales DL y RBF aplicadas a un dataset con información sobre eventos de conducción.

Abstract

Deep learning must be worked offline because the training and execution phases are processed separately. This process often requires different computers due to the requirements to obtain the models. A limitation of this approach is that the training data cannot be continuously incorporated into the knowledge base because a long and complex training process needs to be repeated to obtain the new models. Although the environment may not be static, it is crucial to dynamically train models by integrating new information during execution. Radial Basic Functions Neural Networks allow for continuous learning due to their high learning speed and simple architecture. Thus, the RBFNN allows for a direct interpretation of the function, with each node having only one hidden layer. The study analyzed a dynamic dataset approach by adding new data to the training on the fly, given the constant changes in the driver's data, vehicle information, environmental conditions, and traffic accidents. The dataset used to predict accident risk levels was obtained from previous doctoral and master's works at Escuela Politécnica Nacional that contained information on driving events. Finally, this study compares the processing time and performance of Convolutional Neural Networks (CNN) with several Machine Learning (ML) algorithms, including RBF, Multilayer Perceptron (MLP), and Random Forest (RF), using evaluation metrics of accuracy, specificity, and sensitivity-recall. The results provide recommendations for new accident prediction models.