Home
As part of the introductory course to machine learning, our project focuses on recognizing road signs. Our team is made up of Jade Clouâtre, Isabelle Eysseric, Damien LaRocque and Benoît Verret.
As long as roads are used by vehicles driven by humans, most of the information available on the state of the road or on the directions to follow to reach your destination is given by road signs placed near the roadway. In the near future, when autonomous driving will tend to become more widespread, autonomous cars will therefore need to be able to distinguish between road signs in order to obey the highway code.
Until now, most of the studies on the subject in Canada have been conducted in Ontario[1]-[4]. Given that road signs and road networks are under provincial jurisdiction, these studies are not adapted to the Quebec context, especially since Quebec is the only Canadian province to have unilingual signs in French. The goal of this project is therefore to determine whether the data available in the public domain are reliable enough to produce a model for recognizing Quebec road signs, or whether it will be necessary to establish a database of Quebec road signs for the subsequent development of models adapted to Quebec.
Given that road sign recognition is based on images taken with cameras, it will be necessary to adopt a method that has proven itself in the field of computer vision. Two methods can be considered for this purpose: deep learning and the Viola-Jones method[5]. Although the Viola-Jones method is the least excessive, it addresses problems of localizing objects on an image rather than recognition problems. In doing so, it proves to be useless, because the data it requires must contain location coordinates of the signs on the images, which may be superfluous for our project.
This is why we are considering opting for deep learning, especially since we have a large mass of data available worldwide and we will have to sort the data according to several classes. The library will be used to implement the algorithm.
The training will be done with images from different datasets, prioritizing images of signs in a road environment. Indeed, we want to avoid having a model that is not able to detect a sign in a busy environment, as would be the case if we were training with images on a white background.
Similarly, for the purposes of simplification and speed of training, we will use a transfer learning approach[6]. This approach will first consist of training a convolutional neural network with the largest datasets even if they do not contain Quebec data[7].
With large databases similar to our target, those of the different countries, we will do deep learning of the different layers of our neural network by initializing the weights randomly. Then we will re-train the last layer of our model with our small dataset from Canada. Finally, we will use backpropagation to refine the weights of the other layers by initializing the weights, this time, with those obtained in training on the large dataset.
Given the method considered and the number of classes to be taken into account, it is essential to have enough data to train the model efficiently. Since road signs are standardized internationally by the Vienna Convention on Road Signs and Signals[8], it is possible to rely on several datasets from different countries for the project. A census of databases in the field was carried out. Table 1 lists the datasets that were identified during this census.
Table 1 – List of datasets available for classifying road signs
And finally, this project aimed to determine whether the data available in the public domain are reliable enough to create a model that can recognize Quebec road signs.
The models that were developed and the analyses that were conducted for this purpose showed that the transfer learning performance on a pre-trained model is not satisfactory, particularly given the limited Quebec data that is readily available. It is therefore appropriate to describe a data set based on images taken in the field in order to train a new model. Similarly, before implementing a model of this type in an autonomous vehicle, it would be appropriate to add corresponding classes to the panels that were ignored for this project.
[1] S. Zabihi, "Detection and Recognition of Traffic Signs Inside the Attentional Detection and Recognition of Traffic Signs Inside the Attentional Visual Field of Drivers Visual Field of Drive", Master's thesis, University of Western Ontario, 2017.
[2] J. Feng, "Traffic Sign Detection and Recognition System for Intelligent Vehicles", 2014.
[3] A. Mammeri, A. Boukerche, J. Feng and R. Wang, "North-American speed limit sign detection and recognition for smart cars", in the 38th Annual IEEE Conference on Local Computer Networks - Workshops, 2013, p. 154-161.
[4] R. Mukhometzianov and Y. Wang, "Review. Machine learning techniques for traffic sign detection", CoRR, t. abs/1712.04391, 2017.
[5] P. Viola and M. Jones, "Rapid object detection using a boosted cascade of simple features", in Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, vol. 1, 2001, p. I-I.
[6] S. Chilamkurthy. (2017). "Transfer Learning for Computer Vision Tutorial", available at (visited on 10/19/2020).
[7] E. Stevens, L. Antiga and T. Viehmann, "Deep Learning with PyTorch". Manning Publications, 2020.
[8] United Nations (1968). "Convention de Vienne sur la signalisation routière", available at (visited on 10/19/2020).
[9] Ruhr-Universität Bochum — Institut für Neuroinformatik (INI). (2019). "German Traffic Sign Detection Benchmark", disponible à l' (visited on 10/19/2020).
[10] ——, (2019). "German Traffic Sign Recognition Benchmark", available at (visited on 10/19/2020).
[11] S. Houben, J. Stallkamp, J. Salmen, M. Schlipsing and C. Igel, "Detection of traffic signs in realworld images: The German traffic sign detection benchmark", in the 2013 International Joint Conference on Neural Networks (IJCNN), IEEE, August 2013.
[12] R. Timofte. (2014). "BelgiumTS Dataset", available at (visited on 10/19/2020).
[13] M. Mathias, R. Timofte, R. Benenson and L. Van Gool, "Traffic sign recognition — How far are we from the solution?" in the 2013 International Joint Conference on Neural Networks (IJCNN), IEEE, 2013, p. 1-8.
[14] R. Timofte, K. Zimmermann and L. V. Gool, "Multi-view traffic sign detection, recognition, and 3D localisation", Machine Vision and Applications, t. 25, no 3, p. 633-647, in 2011.
[15] Laboratory for Intelligent and Safe Automobiles - CVRR - UCSD. (2012). "LISA Traffic Sign Dataset", available at (visited on 10/19/2020).
[16] A. Mogelmose, M. M. Trivedi and T. B. Moeslund, "Vision-Based Traffic Sign Detection and Analysis for Intelligent Driver Assistance Systems: Perspectives and Survey", IEEE Transactions on Intelligent Transportation Systems, t. 13, no 4, p. 1484-1497, 2012.
[17] Google Open Source. (2020). "Open Images Dataset V6", available at (visited on 10/19/2020).
[18] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, T. Duerig and V. Ferrari, "The Open Images Dataset V4: Unified image classification, object detection, and visual relationship detection at scale", IJCV, 2020.
[19] L. Huang. (2014). "Chinese Traffic Sign Database", available at (visited on 10/19/2020).
[20] J. Liu, L. Huang and B. Niu, "Road sign text detection from natural scenes", in 2014 International Conference on Information Science, Electronics and Electrical Engineering, vol. 3, 2014, pp. 1547-1551.
[21] Y. Zhang, Z. Wang, Y. Qi, J. Liu and J. Yang, "CTSD: A Dataset for Traffic Sign Recognition in Complex Real-World Images", in IEEE Visual Communications and Image Processing (VCIP), 2018, pp. 1-4.
[22] GéoInformations - Espace interministériel de l’information géographique. (2013). "Panneaux routiers au format SVG et PNG par Bertrand BOUTEILLES", available at (visited on 10/19/2020).
[23] Ville de Montréal. (2019). "Signalisation routière (excluant le stationnement)", available at (visited on 10/19/2020).
[24] Ministère des Transports du Québec. (2013). "Répertoire des dispositifs de signalisation routière du Québec", available at (visited on 10/19/2020).
[25] Mapillary. (2020). "Mapillary Vistas Dataset", available at (visited on 10/19/2020).