Although pedestrian facility data can improve pedestrian safety, it is not widely available at state level. Challenges such as high cost for time and labor inherent in current data collection methods greatly impede the progress of collecting pedestrian facility data at a state level. This project developed an innovative system to apply image processing and deep learning methods to automatically collect major pedestrian facility data, including sidewalk presence, crosswalk presence, and crosswalk length, from aerial images. Using OpenStreetMap pedestrian facility location data and images from Bing Maps, four core models integrated in the system were developed (using convolutional neural networks based on the VGG16 architecture) to prepare model development data (50,112 images), detect presence of facilities (with an accuracy of 96.35% ~ 98.43% for aerial crosswalk images, and 92.87% for aerial sidewalk images), check the ground truth when street view data are available, and measure the length of the detected crosswalk. The aerial image provided by the Mississippi Department of Transportation was tested to evaluate the validity of the system on unseen data. In the test, 400 images from Forrest County, Mississippi were processed, resulting in an accuracy as high as 99.23% for crosswalk detection, 91.26% for sidewalk detection, and 93.7% for crosswalk length mensuration. A system was developed to process aerial images of candidate locations. Final outputs are then stored in an excel table. Using the system, the time cost to collect data from an area of 466.31 square miles would be approximately 8 days on a GPU workstation pc. This automated data collection method has potential to greatly decrease the monetary and time cost of state-level pedestrian facility data collection and provide the foundation for the next generation data collection method.

The final report is available.