Mexican traffic sign detection and classification using deep learning

Automatic detection and classification of traffic signs is challenging to support a driver’s safety and even assist in autonomous driving. This paper aims to propose a methodology for detecting and classifying Mexican traffic signs using deep learning. The methodology consists of the creation of a new Mexican traffic sign data set, the training, testing, and comparing of two sign detectors (the Region-based Convolutional Neural Network (R-CNN) and the You Only Look Once (YOLO v3)), and the use of a modified Residual Neural Network (ResNet-50) for classification. According to the detection results, the combination R-CNN/ResNet-50 yielded a mean Average Precision (mAP) of 95.33%, while the YOLO v3/ResNet-50 yielded 90.33%. The overall classification accuracy was 99.00%. Our results are competitive to those presented in the literature. We demonstrated the robustness of our proposal by conducting a test to classify images that do not contain traffic signs. The accuracy for the R-CNN/ResNet-50 was 99.5% and 99.77% for the YOLO v3/ResNet-50. We also obtained satisfactory classification results with traffic signs occluded and inserted in random positions in the scene. Finally, an ablation study regarding the data set and the batch size was conducted. •A new data set with 1426 Mexican traffic signs is presented.•We compare an R-CNN with YoloV3 for the stage of traffic sign detection.•The stage of image classification is implemented using a modified ResNet-50.•We present tests with traffic signs occluded and randomly inserted in the scene.•The results obtained are in light of the obtained in the state-of-the-art.