Deep-learning-based Automatic Detection and Classification of Traffic Signs Using Images Collected by Mobile Mapping Systems

As interest in autonomous driving has increased in recent years, various sensors have been developed for use in vehicles to detect and classify traffic signs. When road traffic facilities are not recognized owing to the malfunction of sensors, point cloud data and images collected by mobile mapping systems (MMSs) are used to construct high-definition maps containing road traffic facility information. However, when traffic signs, among the targets of high-definition map construction, are constructed using point cloud data, it becomes difficult to detect and classify traffic signs because they are highly reflective. In this study, we detected and sub-classified traffic signs by combining Mask Regions with Convolutional Neuron Network (Mask R-CNN) and Inception-v3 models based on image data obtained using MMSs. Image data obtained by various types of MMS were used to detect traffic signs and classification results were verified. The detection accuracy of traffic signs was 87.6% and the classification accuracy was 77.5%; thus, the method proposed in this study can be used to automatically construct traffic signs for high-definition maps.