Traffic sign integrity analysis using deep learning

Standardization and harmonization of high-quantity and low-cost roadway assets such as traffic signs make a compelling case now, especially in the matter of road safety and the advent of autonomous transportation vehicles. As for the standardization and degradation problem in classifying traffic signs, no automatic approach has been developed yet. This paper presents two trained models that can classify the compliance of a traffic sign to a prescribed design standard in accordance to the Vienna Convention on Road Signs and Signals adapted by the Philippines, and its physical degradation status. Due to the limited size of image dataset and a weak computing resource, a deep learning approach using transfer learning was utilized. The process highlights the use of a state-of-the-art pre-trained convolutional neural network - the ResNet-50 which was trained on a subset of the ImageNet database. This network was then fine-tuned on the curated dataset of traffic signs derived on Google Street View with varying design and physical condition. The experiments conducted exhibit test accuracy of 95.70% in the compliance model and 96.20% in the degradation model which shows a great potential in developing a streamlined traffic sign assessment platform.