An Automated System for Pedestrian Facility Data Collection from Aerial Images

This project will develop a novel data collection system to automatically detect, classify, and measure major pedestrian facilities (e.g., sidewalks and crosswalks) from aerial images. Work in Stage 1 that focuses on developing functional models to automatically acquire labeled aerial images, train the facility detection model, check ground truth for occluded facilities, and measure numerical features of crosswalks. A model will be developed that can automatically prepare labeled sample data, for example, as “having a sidewalk” or “not having a sidewalk.” A specific number of aerial images labeled as “sidewalk”, “crosswalk”, or “none” will be collected and divided for training and validation. Input data in uniform scale and quality will be generated before training by augmenting sample data prepared by the model. A deep-learning-based model will be built and its various architectures will be evaluated. The most efficient one will be chosen to detect whether a facility is “present”, “absent”, or “occluded” in a given aerial image. A model will be built to classify partially or completely occluded facilities not fully detectable in the previous model. It infers the occluded areas with multiple priors, which fuse the knowledge about roads, occluders, facility structures, and especially, street view images. Once a facility is detected as a “marked crosswalk”, the length of the crosswalk edge will be measured automatically from the aerial image using computer image processing techniques. For occluded crosswalks, assumptions will be made by fusing the research team’s knowledge about road and crosswalk structure. In Stage 2, the data collection system will be developed by connecting and integrating the four models. Performance evaluation will then be conducted to assess accuracy of the detected data and the system’s efficiency. Stage 2 will also involve testing with aerial images from Caltrans and Mississippi DOT to collect adjustments and feedback from state agencies. The system will be completed by connecting and integrating the aforementioned functional models and its performance assessed by comparing the data collected by the proposed system and Caltrans’ records. A summary of potential adjustments will be documented for future implementation and improvement. A draft final report will be prepared including all relevant data, methods, models, and conclusions along with guidance on how to use the system to collect pedestrian facility data in a state DOT.

Language

  • English

Project

  • Status: Active
  • Funding: $129972
  • Contract Numbers:

    Project 20-30, IDEA 209

  • Sponsor Organizations:

    National Cooperative Highway Research Program

    Transportation Research Board
    500 Fifth Street, NW
    Washington, DC  United States  20001

    American Association of State Highway and Transportation Officials (AASHTO)

    444 North Capitol Street, NW
    Washington, DC  United States  20001

    Federal Highway Administration

    1200 New Jersey Avenue, SE
    Washington, DC  United States  20590
  • Project Managers:

    Jawed, Inam

  • Performing Organizations:

    University of Southern Mississippi

    ,    
  • Principal Investigators:

    Zhang, Yuanyuan

  • Start Date: 20190408
  • Expected Completion Date: 0
  • Actual Completion Date: 0

Subject/Index Terms

Filing Info

  • Accession Number: 01701575
  • Record Type: Research project
  • Source Agency: Transportation Research Board
  • Contract Numbers: Project 20-30, IDEA 209
  • Files: TRB, RiP
  • Created Date: Apr 8 2019 3:06PM