September 2023

53 52 Artificial intelligence, or AI, is an exciting field that I’m sure many of you have started to hear about with increasing frequency in the media! From ChatGPT to artgenerating software to images of robots taking over the world, AI can often seem scary and futuristic without a proper understanding of this complex field. Obviously, AI has some pretty complicated implications for society that I can’t even begin to unravel or even imagine at the moment, but what I find most interesting is how AI can be used to help with research and advancing scientific data gathering. For example, this summer, I worked with and trained AI to detect crosswalks in Florida, specifically in Hillsborough County. I wanted to know the impact of faded pavement markings on accidents in residential areas primarily occupied by White vs Latino populations. I wanted to know if faded crosswalks were more likely to occur in minority neighborhoods in order to inform policy regarding government funding for transportation signage. My research was done at the Resilient Infrastructure and Disaster Response Center (RIDER). This is a shared research institute at Florida Agricultural and Mechanical University and Florida State University. I worked under the mentorship of graduate students R. Antwi, S. Takyi, and Dr. M. Koloushani. The head of RIDER is Dr. E Ozguven, who also added supervision to this project. They had completed research regarding training the AI to detect crosswalks and used it towards conducting research on speeding in school zones. The first step was to use aerial imaging to view all the crosswalks in my study area. You can see how AI might be helpful for this task, as obviously, it would be quite difficult for me to count thousands of crosswalks in Hillsborough! In addition, how could I measure if a crosswalk was faded and define what this meant? To use AI, you first need to train it to recognize images by repeatedly feeding it examples of what you want it to identify. In this vein, we used the YouOnly-Look-Once (YOLOv2) object detection algorithm coupled with Geographic Information System (GIS) techniques to limit the dataset, establish a geocoded connection for each detected crosswalk, and conduct a sensitivity analysis to identify different types of crosswalks and their locations. The proposed approach can inventory the crosswalks with 85.9% recall and 88.7% precision, utilizing data from all 67 counties in Florida. RIDER went to great lengths to use accurate and copious amounts of data; as for this study, data from the most recent images (as of December 2019) were obtained with a total size of 1.2 terabytes. Utilizing GIS, we could geographically place the crosswalks by importing the images into a mosaic dataset within the GIS software. Mosaic datasets manage and display multiple geocoded images, allowing the AI to geolocate imaging it detected as crosswalks. Then, the YOLOv2 object detection algorithm broke down the images we fed into pixels and searched the GIS map for similar images, providing a confidence level each time. The confidence level ranged from 0-1, with 1 meaning the AI is completely sure that it located a crosswalk, and this confidence level is highly proportional to the fading of the crosswalk. This data is extremely powerful because it allowed us to gain vast knowledge and insights into Florida crosswalk locations and their effect on pedestrian crashes. In addition, using the fading level allows for uncovering systemic bias, as I was able to prove that there is an initial correlation between faded crosswalks and Hispanicmajority neighborhoods. While AI might seem scary initially, it’s important to remember that it can be humanity’s greatest helper when used correctly. Using AI To Gain Social Equit y By Hannah Flitman, Grade 11, Berkeley Preparatory School, Tampa Stock picture by Uriel Mont from Pexels