Geolocating and Grading Crosswalks using Deep Learning

Many Departments of Transportation (DOTs) in the United States have incomplete accounting of the location and condition of their crosswalks. Deploying inspection crews to inspect crosswalk conditions is costly and time-consuming. A complete inventory of crosswalks and their associated paint condition would allow maintenance cycle optimization and reduce inspection costs. Using aerial imagery from the Washington, D.C. metropolitan area, a Mask R-CNN was trained to identify, measure, and localize unique object instances of crosswalks. These crosswalks will serve as the reference dataset for ongoing crosswalk quality assessment. A similar network was trained to identify crosswalk instances in consumer dashcam imagery. Crosswalks extracted from these images are used to update the quality assessment of the reference dataset. We report the data collection process, approaches to paint quality estimation, and deep learning implementations. Crosswalks were located and cataloged using aerial imagery with a precision of 0.968 and a recall of 0.949 within a test region. A quantitative paint quality metric was successfully established and calculated from street-level imagery. By attributing these scores to localized crosswalks, we are able to automatically provide city-wide paint quality assessments. This method serves as a promising technique for DOTs to continuously monitor the condition of their crosswalks and spare valuable department resources.