Deep learning methods for recognizing signs/objects in road traffic
Josef Bengtson, Filip Heikkilä, Per Nilsson, Lukas Nyström, Erik Persson, Gustav Tellwe · Chalmers Publication Library (Chalmers University of Technology) · 2018
In this thesis the authors have used three different neural network structures, Convolutional Neural Network (CNN), Capsule Network and Faster R-CNN, to detect and classify traffic signs from the German Traffic Sign Recognition-and Detection Benchmark (abbreviated GTSRB and GTSDB).The networks were implemented using machine learning libraries for Python such as TensorFlow and Keras.The results obtained were a classification accuracy of 98.61 % for the CNN and 99.62 % for the Capsule Network.The aim was to attain results close to the results of the best network listed at GTSRB, which is 99.71%.Another goal was to use a deep learning networks to locate and classify a traffic sign within an image.This was done with an implementation of a Faster R-CNN, which was able to successfully detect and classify signs with an mAP of 0.56.It was noted that the network was able to detect most of the signs, however it struggled with classification.For this reason a version that combined Faster R-CNN with a Convolutional Neural Network trained on GTSRB was created, which outperformed the Faster R-CNN.