A Data Driven Approach to Creating a Traffic Sign Asset Inventory using Remote Sensing Technology

• Author / Creator

  Karsten, Lloyd Andrew Maurice

• Traffic signs play a critical role in the safety and efficiency of any roadway, but with limited information on the current status of traffic sign inventories, the placement and condition of traffic signs goes largely unchecked. Therefore, the collection of a complete traffic sign inventory (TSI) is needed to ensure traffic signs meet the needs of the current driving population. However, the size of current global traffic sign networks makes applying traditional survey methods to the collection of a TSI difficult, if not economically infeasible. Therefore, there is room for technological and methodological advancement to create a new TSI process to inventory and analyze traffic signs. This thesis proposes the use of light detection and ranging (LiDAR) and video-log imaging to conduct an automated extraction of a TSI. The details of traffic sign location, orientation, placement, and panel classification define the fundamental components of a TSI.Traffic signs are extracted through a Gaussian mixture model, density clustered, and filtered for flatness before measuring the vertical and horizontal orientation through principal component analyses. Traffic sign placement is dependent on the location of lane markings. Therefore, the road surface near each traffic sign is extracted, rasterized, and intensity manipulated to determine the linear lane marking intensity edges. The markings are used to determine the lateral and vertical placement of the traffic signs. Sign classification is determined from video-log images by applying a trained GoogLeNet convolutional neural network. This completes the traditional TSI and creates a platform with which to analyze the efficacy of traffic sign placements. Additionally, the detail provided by LiDAR scans allows for the measurement of the visibility of the traffic signs, which is unavailable through traditional surveying methods. This is used to assess the time available to drivers to read and react to traffic signs placed along the segment. A 4-km test segment was utilized to assess the accuracy of the proposed method, providing an Eastbound TSI for 30 traffic signs.The intensity-based extraction of traffic signs had a precision, recall, and F1-Score of 98.3%, 92.06%, and 95.08%, respectively. The extraction of lane markings had a precision, recall, and F1-Score for the left-lane markings of 100%, 89.36%, and 94.38%, respectively. The corresponding values for right-lane markings were estimated as 93.47%, 86%, and 89.58%, respectively. The image classifier had a sample of 13,604 training images spanning 155 traffic sign classes and 10 false positive classes within Alberta. The structure is trained within a half hour on GPU and ~8 hours on CPU and produces 83.6% accuracy on the validation set. This translates to a Top-1 and Top-2 classification error of 10.35% and 3.24%, respectively. However, when applied to the original video-log images, the sliding window procedure used to apply the trained classifier to cropped image samples creates the opportunity for misclassifications across the video-log image. This reduces the accuracy of the classifier to 53.3%. Finally, the visibility assessment considered day and night-time conditions as well as the impact of consecutive placement (i.e. driver’s attention fixated on only the nearest sign). This provides a discussion of how available visibility affects different driving populations and which traffic signs are most susceptible to being missed while driving.This thesis presents a method for the expedited accurate extraction of a TSI, including location, orientation, placement, classification, and visibility. Utilizing the detail available from high-density LiDAR scans, the extractions are completed with a high degree of accuracy and with time benefits over the traditional manual methods. The contributions of this thesis include (i) proposing a method for the extraction of a TSI, (ii) assessing sign visibility, and (iii) creating Canada’s first traffic sign image database. This sets the stage for continued research into the extraction of TSIs, the continued development of a traffic sign image database for Canada, and guidance for industry professionals considering using LiDAR or video-logs for creating a TSI.

• Subjects / Keywords

  • Asset Management
  • Traffic Sign
  • LiDAR
  • Remote Sensing
  • Inventory

• Graduation date

  Spring 2019

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-7c0y-k709

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.