An Evaluation of the Machine Readability of Traffic Sign Pictograms using Synthetic Data Sets

We compare the machine readability of pictograms found on Austrian and German traffic signs. To that end, we train classification models on synthetic data sets and evaluate their classification accuracy in a controlled setting. In particular, we focus on differences between currently deployed pictograms in the two countries, and a set of new pictograms designed to increase human readability. We find that machinelearning models generalize poorly to data sets with pictogram designs they have not been trained on, and conclude that manufacturers of advanced driver-assistance systems (ADAS) must take special care to properly address small visual differences between different traffic sign pictogram designs. Our main contributions are the creation of a vast synthetic data set of traffic sign images, training and evaluating state-of-theart classification models to assess the machine readability of different pictogram designs, and employing techniques from explainable AI to analyze which image regions are particularly important to the classifiers.