Member-only story
GeoAI : Automating Pedestrian Crosswalk Investigation and Assessing Marked Crosswalk Quality Using MaskRCNN
Urban Transportation and Pedestrian Crosswalk
Introduction
Crosswalks are the locations where we expect pedestrians to crossroads and streets. Marked pedestrian crossings are often found at intersections, but may also be at other points on busy roads that would otherwise be too unsafe to cross without assistance due to vehicle numbers, speed or road widths. There are a variety of different pedestrian crossings with different meanings and rules. High-quality marked crosswalks support a safe walkable urban environment. Drivers tend to stay out of pedestrian crosswalks if the crosswalks are delineated with paint striping. Painted crosswalks serve as another reminder for motorists to stop clear of the intersection. They also encourage pedestrians to cross at the intersection rather than risk a mid-block crossing. There are plenty of crosswalks that are unmarked or poorly maintained in modern cities. The painted area of a crosswalk is often considered the legal boundary of the crosswalk, so it can have legal implications as well (especially when a traffic accident happens). Being the largest city in Canada, pedestrian safety is the main concern of everyone living in Toronto. Pedestrians are the main victims of road collisions in Toronto. So it is important to inspect pedestrian crosswalks and to assess marked crosswalks quality on a continual basis to enhance pedestrian safety. In my previous posts, I covered how to use Fastai SSD to detect objects. Here I would like to share one of my previous Deep Learning work samples about automating pedestrian crosswalk investigation using MaskRCNN and evaluating crosswalk quality. Let’s get started by preparing training sample data.
Data Preparation: Create classes and label objects
All supervised Deep Learning tasks depend on labelled datasets. Image annotation, or labelling, is vital for deep learning tasks such as computer vision and learning. Creating crosswalk sample data is easier if it is based on intersection data.
