Date of Award

Summer 8-2022

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

School

Computing Sciences and Computer Engineering

Committee Chair

Dr. Chaoyang Zhang

Committee Chair School

Computing Sciences and Computer Engineering

Committee Member 2

Dr. Yuanyuan Zhang

Committee Member 2 School

Leadership

Committee Member 3

Dr. Zhaoxian Zhou

Committee Member 3 School

Computing Sciences and Computer Engineering

Committee Member 4

Dr. Bo Li

Committee Member 4 School

Computing Sciences and Computer Engineering

Committee Member 5

Dr. James V. Lambers

Committee Member 5 School

Mathematics and Natural Sciences

Abstract

Large-scale collection of pedestrian facility (crosswalks, sidewalks, etc.) presence data is vital to the success of efforts to improve pedestrian facility management, safety analysis, and road network planning. However, this kind of data is typically not available on a large scale due to the high labor and time costs that are the result of relying on manual data collection methods. Therefore, methods for automating this process using techniques such as machine learning are currently being explored by researchers. In our work, we mainly focus on machine learning methods for the detection of crosswalks and sidewalks from both aerial and street-view imagery. We test data from these two viewpoints individually and with an ensemble method that we refer to as our “dual-perspective prediction model”. In order to obtain this data, we developed a data collection pipeline that combines crowdsourced pedestrian facility location data with aerial and street-view imagery from Bing Maps. In addition to the Convolutional Neural Network used to perform pedestrian facility detection using this data, we also trained a segmentation network to measure the length and width of crosswalks from aerial images. In our tests with a dual-perspective image dataset that was heavily occluded in the aerial view but relatively clear in the street view, our dual-perspective prediction model was able to increase prediction accuracy, recall, and precision by 49%, 383%, and 15%, respectively (compared to using a single perspective model based on only aerial view images). In our tests with satellite imagery provided by the Mississippi Department of Transportation, we were able to achieve accuracies as high as 99.23%, 91.26%, and 93.7% for aerial crosswalk detection, aerial sidewalk detection, and aerial crosswalk mensuration, respectively. The final system that we developed packages all of our machine learning models into an easy-to-use system that enables users to process large batches of imagery or examine individual images in a directory using a graphical interface. Our data collection and filtering guidelines can also be used to guide future research in this area by establishing standards for data quality and labelling.

ORCID ID

0000-0001-5832-4487

Share

COinS