The aim of the project was to construct a system that recognizes road signs in photographs. In this project, we created a classical learning system that based on prior knowledge for classification. The chosen classifing algorithm is Support Vector Machine (SVM).

The problem:

A road sign recognition system faces a classical problem of pattern recognition, meaning classifing between different road signs. On top of that, the location of the road sign in the picture is unknown. Once these obstacles are overcome, such system could be integrated in a Smart Driver System. 

The solution:

The suggested method for candidate objects detection was color segmentation. In the process, objects which aren’t road signs are located as well. These objects are filtered later during the classification stage.
Each object goes through image processing, and is represented in the feature space. It is then labeled to its category, according to the system mode. Labeling is done either by the user while building the database, or by the system call during classification.

The chosen classifying algorithm is Support Vector Machine (SVM). The algorithm was trained to the optimal classification model, based on the database. New objects are classified to different road signs by the optimal model. 

In this project, the classifier is composed of a three phase decision tree. The first phase distinguish road signs from other objects. It is followed by assigning the road sign to groups by shape and color  in the second phase. The last phase does the fine classification between road signs in their group.

Figure 3 - Classification decision tree

In order to make our program user friendly, Graphical User Interface was designed and implemented. 

Figure 5 - Classification GUI demonstration

The project was programmed in Matlab 6.5, on PC platform. Our software communicates with MS-Access and LIBSVM. Photographs were taken by VISL’s digital camera using a standart 640x480 resolution.
 

Conclusions
In this project, a road sign recognition system was implemeted. 1700 records database was built from 400 pictures of 14 different road signs. The classification process contains three phases. The system performance was estimated for each classification phase separately by Cross Validation. The optimal model was found to have low error rates in each stage.

Acknowledgment

We are grateful to our project supervisor Dori Peleg for his help and guidance throughout this work. We would also like to thank the lab staff, Johanan Erez and Ina Krinsky for their technical support.
We are also grateful to the Ollendorff Minerva Center Fund for supporting this project.

FULL DOCUMENTATION