Middle-East Traffic Sign (METS) recognition: databases and models

Abstract

Automatic Traffic Sign Recognition is a crucial sub-task for autonomous driving, playing a key role in reducing accidents and enabling safer, more intelligent navigation on the roads. This challenge remains largely unresolved in the Arabic world, unlike other regions. This paper reviews the current state-of-the-art in automatic Arabic road-signs recognition; it proposes a new model, which improves the best performance of existing databases; it merges the Arabic Traffic Signs collection with the Jordanian dataset, and it uses several data augmentation techniques to create the largest publicly available database in the Arabic world to date. Finally, it gives a benchmark algorithm for the newly introduced collection. The proposed model is a variation of the lightweight Residual Net architecture, which has been improved by a significant reduction in the number of parameters, minimization of down-sampling to maintain as much spatial detail as possible, and the incorporation of both Dropout and Batch Normalization to ensure better regularization and overfitting prevention. The performance of the introduced model on the Arabic Traffic Sign database surpasses the current state of the art in both accuracy and training time. The same architecture is also very successful when tested on the newly introduced picture collection.