Reliable Traffic Sign Recognition: Are We Finally There?

Traffic Signs Recognition (TSR) is a fundamental task for the implementation of automated driving assistance systems. Major technical advancements have been achieved thanks to the adoption of machine learning classifiers, and especially Deep Neural Networks (DNNs), which were successfully used to correctly classify most of the traffic signs. A small number of misclassifications persist, that are due to various reasons ranging from a flawed learning process of classifiers to low-quality images, adverse environmental conditions, or adversarial attacks to the TSR. Misclassifications are unlikely, but still, they are a severe safety hazard. This paper discusses five strategies that can be adopted for safe, robust, and reliable TSR, that comprise data augmentation, noise removals, corruption, and classifier misclassification detection. We compare these strategies in an experimental campaign, where we inject camera failures into images of traffic signs from two public datasets. Results show that corruption and misclassification detectors can identify and discard in advance a huge fraction of the images that are going to be misclassified but may as well discard images that would have been correctly classified. We conclude that corruption and misclassification detectors can be used with DNNs to reduce misclassifications of TSR components and provide a more reliable functionality.