121. DETECTION OF US TRAFFIC SIGNS USING COMPUTER VISION

Department: Electrical & Computer Engineering
Faculty Advisor(s): Mohan Trivedi | Thomas Moeslund

Primary Student
Name: Andreas Moegelmose
Email: amoegelmose@ucsd.edu
Phone: 858-822-0002
Grad Year: 2012

Abstract
This work is on detecting American traffic signs using computer vision. Recently, a number of high-end car manufacturers have been introducing speed limit sign detection in the top of the range models. The detection, however, only works for speed limit signs in Europe, which are circular with a bright red border. Very little work on detecting American signs has been published, and very little work has been done in properly integrating sign detection with a driver-in-the-loop driver assistance system. This work investigates the current state-of-the art in traffic sign detection and attempts to apply the current best algorithms to American signs. Necessary alterations for American signs are made and the performance of the algorithms is compared. Some of the most used algorithms rely on training with images of real signs. This work also attempts to substitute real-life training data with synthetic training data in order to make training of theses systems easier. The performance of synthetic training data is compared to the performance of real-life training data and it is determined whether using synthetic training data in the training of traffic sign detection systems is feasible.

Related Links:

  1. http://cvrr.ucsd.edu/
  2. http://cvrr.ucsd.edu/lisa/

« Back to Posters or Search Results