Deep Learning Applications for Traffic Sign Detection and Classification

The paper addresses a new method proposed for detecting and identifying road infrastructure elements. The method can also be used for other image processing and analysis tasks. The method uses a YOLO-type single-stage visual detector to predict the coordinates of a certain number of bounding boxes with a MobileNet convolutional neural net used as a classifier. The tracking mechanism makes it possible to link the frames by assigning a unique number to every detected object. A large set of 160 000 traffic signs was used as the training data set for the model. A data flow architecture has been developed. The metrics generated and the rate of inference are sufficient to use as a model for collecting new data. The findings can be used for further analysis. The system is already being used in a road structure inventory project.