Application of Machine Learning Techniques for Predicting Potential Vehicle-to-Pedestrian Collisions in Virtual Reality Scenarios

Abstract

The definition of pedestrian behavior when crossing the street and facing potential collision situations is crucial for the design of new Autonomous Emergency Braking systems (AEB) in commercial vehicles. To this end, this article proposes the generation of classification models through the deployment of machine learning techniques that can predict whether there will be a collision depending on the type of reaction, the lane where it occurs, the visual acuity the level of attention, and consider the most relevant factors that determine the cognitive and movement characteristics of pedestrians. Thereby, the inclusion of this type of model in the decision-making algorithm of the AEB system allows for modulating its response. For this purpose, relevant information on pedestrian behavior is obtained through experiments made in an ad-hoc, Virtual Reality (VR) environment, using a portable backpack system in three urban scenarios with different characteristics. Database generation, feature selection, and k-fold cross-validation generate the inputs to the supervised learning models. A subsequent analysis of the accuracy, optimization, error measurement, variable importance, and classification capability is conducted. The tree-based models provide more balanced results for the performance metrics (with higher accuracy for the single decision tree case) and are more easily interpretable and adaptable to the algorithm. From them it is deduced the high importance of the reaction type and the relative position where it occurs, coinciding with the high significance of these factors in the analyzed collisions.