A Two-stage Learning Approach for Traffic Sign Detection and Recognition

With the progress of advanced driver assistance systems (ADAS), the development of assisted driving technologies is becoming more and more important for vehicle subsystems. The traffic signs are designed to remind the drivers of possible situations and road conditions to avoid traffic accidents. This paper presents a two-stage network to detect and recognize the traffic sign images captured by the vehicle on-board camera. In the detection network, we adopt Faster R-CNN to detect the location of the traffic signs. For the classification network, we use SVM, VGG, and ResNet for validation and testing. We compare the results and integrate the detection and classification systems. The datasets used in this work include TT100K and our own collected Taiwan road scene images. Our technique is tested using the videos acquired from the highway, suburb and urban scenarios. The results using Faster R-CNN for detection combined with VGG17 for classification have demonstrated superior performance compared to YOLOv3 and Mask R-CNN.