Detection of curbside storm drain from street level images using Faster R-CNN
Stormwater management is a significant part of modern urban infrastructure. With an increase in climate change due to global warming, this system plays a major role in conserving water and maintaining the environment. Also, they play a significant role in risk management in times of flood. Storm Drains/inlets are essential in modeling this system. Precise mapping of the location of these drains is the key step in improving the infrastructure. State of the art deep learning technique using Faster R-CNN is presented in this thesis to identify the drains on the curb of the road using Google street view images as the primary source. The model is evaluated with 1000 street-level images of streets and highways of Urbana-Champaign, Illinois-USA. The method proposed shows a significant improvement in the detection accuracy of drains by eliminating a significant amount of false positives compared to the previous state of the art machine vision detection techniques. The dataset used for the thesis is available for future researchers.