Traffic Sign Detection with Convolutional Neural Networks

Abstract

This research focuses on improving traffic sign detection in cars using Convoluted Neural Networks (CNN), with images from the German Traffic Sign database. In order to generate more accurate detection results of traffic signs, different algorithms were used to generate the detection and classification tasks. The Faster Region Based Convolutional Neural Network (Faster R-CNN) and You Only Look Once networks were compared beforehand to determine which CNN to use. The Faster R-CNN was decided upon based off of previous results, then used to generate the classification and detection tasks. Pre-training weights were made using Caffe based off of the German Traffic Sign Recognition Benchmark database. Different methods of generation of training data were then used and compared. The Faster R-CNN network was used to create a classification task based off the images from the self-generated training images, which was tested against the German Traffic Sign Detection Benchmark database.