Road obstacle detection by AutoEncoder using vehicle driving information

Abstract

In this paper, we present a study on road obstacle detection using an autoencoder with vehicle driving information. We describe a method for detecting the occurrence and location of a road obstacle using the autoencoder, a machine learning algorithm that aggregates vehicle driving information measured by the electronic toll collection system 2.0 on-board units installed in vehicles as probe data via intelligent transport systems spots. The autoencoder continuously builds a model by learning information on vehicle behavior in a normal traffic flow before the occurrence of a road obstacle, and it detects the ofstacle of it when the output from the model shows a poor fit. This approach is highly applicable to ever-changing traffic flows and to a variety of roadway environments. By computer simulations, we show that the detection method using the autoencoder outperforms the supervised learning method using a support vector classifier.