Multi-class Traffic Sign Detection and Classification Using Google Street View Images

Maintaining an up-to-date record of the location and condition of high-quantity low-cost roadway assets such as traffic signs, is critical to the safety of transportation systems. Despite their importance, today’s video-based data collection and analysis practices are still costly, prone to error, and performed intermittently. While, databases such as Google Street View (GSV) contain street-level panoramic images of all traffic signs and are updated regularly, their potential for creating a comprehensive inventory has not been fully explored. The key benefit of these databases is that once roadway assets are detected, accurate geographic coordinates of the detected assets can be automatically determined and visualized within the same platform. Nevertheless, detecting and classifying roadway assets from GSV imagery is challenging due to their interclass variability and particularly changes in illumination, occlusion, and orientation. This paper evaluates the application of a computer vision method for multi-class traffic sign detection and classification from GSV images. The method extracts images using the Google Street View API and leverages a sliding window mechanism to detect potential candidates for traffic signs. For each candidate, a Histogram of Oriented Gradients (HOG) is formed and is concatenated with a Color Histogram. The HOG+Color descriptors are then fed into multiple one-vs.-all Support Vector Machine classifiers to detect traffic signs and classify them into their specific categories. The experimental results with an average accuracy of 95.5% demonstrate the potential of leveraging GSV images as a viable solution for creating up-to-date inventories of traffic signs.

  • Supplemental Notes:
    • This paper was sponsored by TRB committee ABJ50 Information Systems and Technology.
  • Corporate Authors:

    Transportation Research Board

    500 Fifth Street, NW
    Washington, DC  United States  20001
  • Authors:
    • Balali, Vahid
    • Depwe, Elizabeth
    • Golparvar-Fard, Mani
  • Conference:
  • Date: 2015

Language

  • English

Media Info

  • Media Type: Digital/other
  • Features: Figures; References;
  • Pagination: 14p
  • Monograph Title: TRB 94th Annual Meeting Compendium of Papers

Subject/Index Terms

Filing Info

  • Accession Number: 01552883
  • Record Type: Publication
  • Report/Paper Numbers: 15-3421
  • Files: TRIS, TRB, ATRI
  • Created Date: Feb 5 2015 1:08PM