Roadside Traffic Sign Detection Based on Faster R-CNN

Abstract

This paper presents an improved traffic sign detection method based on Faster R-CNN with dataset augmentation and subcategory detection scheme. Firstly, we extract natural scene frames from given videos and determine 20 categories of traffic signs. Secondly, we extend the image dataset and extract regions of interest, then manually annotate all categories. Thirdly, we train the Faster R-CNN model based on TensorFlow, then test the model and obtain the following evaluation indexes: the mean average precision is 99.07%, the recall rate is 99.66%, and the precision rate is 97.54%. Finally, we add the subcategory detection scheme to determine traffic light states, and we get the following evaluation indexes: the mean average precision is 99.50\%, the recall rate is 100%, and the precision rate is 94.40\%. Our experiments prove the robustness and accuracy for both traffic sign detection and subcategory detection of traffic light.