Exploring semi-supervised learning for deciding the order of vehicles in rear-end crashes

Abstract

Knowing the order of vehicles in rear-end crashes is an important step for further analysis to work towards the goal of reducing fatalities and severe injuries in traffic down to zero. For example, by knowing the order of vehicles a more accurate estimation of the driver’s exposure could be done. The main task of this project was to create a system which could decide the order of vehicles involved in each rear-end crash from the text description in a crash database. To try and address this task semi-supervised learning was chosen as the method. The reason for this choice was because all the crash cases were unlabelled from the beginning and to label all the cases manually would be infeasible but to label a few hundred cases was manageable. Two neural networks were built and tested. The text descriptions of the rear-end crashes from the national crash database in Sweden were transformed to vectors and used as input for the neural networks. The performance of each of the networks was measured. The results suggest that further work are needed before the networks could be used for making reliable decisions, however the experiments give some indications on what parameters should be considered carefully while using semisupervised learning and which parameters that need more exploration.