Skip to content

Flag-cs zobrazit českou verzi


Year: 2013
Author: Petr Zimmermann, Ivan Pirner, Marek Hrúz, Miloš Železný, Jan Vodňanský, Jan Kufner

Traffic signs detection and classification

The classifier is designed to detect and classify czech traffic signs. It consists of several modules, which are implemented in the programming language C ++ using OpenCV 2.4.5. The output informations are used for passportization. 


Classifier scheme


Description of the key Modules

  • Back Projection - Back Projection image is calculated using colored histogram obtained from the training set. The histogram is composed of H and S channel (HSV representation of image) and modified by CLAHE technique (Contrast Limited Adaptative Histogram Equalization).



  • MSER Detection - MSER (Maximally Stable Extremal Regions) technique is used to find most stable regions in image.


  • Classification - Our classification has two phases:

Signs/non-signs classification - The feature vectors of regions are obtained by HOG (Histogram of Oriented Gradients)  descriptor and divided into two classes using SVM (Support Vector Machine) classifier.

Classification of traffic signs - Each traffic sign is assigned to a specific class using a Convolutional Neural Network.

  • Geographic location estimation - A GPX format track log is created by a GPS/GNSS device during recording the video from the vehicle. This log records 10 positions of the vehicle per second and is sychronized with the video track using either a spatial or time stamp. The position is linearly interpolated so that each video frame has its own geographical location. Furthermore the distance and azimuth of the sign found in the image is computed to get a more precise information about the sign location.


For more information, please contact:

Marek Hrúz

Univerzitní 8, 301 00 Plzeň

email:, tel: +420 377 632 555

Miloš Železný

Univerzitní 8, 301 00 Plzeň

email:, tel: +420 377 632 548

Confirmation of usage

Traffic signs classifier was developed under project TA02030673 - DOPANAR. The classifier has been piloted in the CEDA a.s. (CENTRAL EUROPEAN DATA AGENCY), where it is operated as a pilot plant.