Accelerated Disaster Reconnaissance Using Automatic Traffic Sign Detection with UAV and AI
Yichang James Tsai, Wei Cheng · 2019
Because of frequent extreme weather conditions, accelerated disaster reconnaissance has become extremely important. In particular, surveying traffic sign damage and conditions has become essential for determining and prioritizing necessary repair/replacement. Under the research project sponsored by the National Academy of Sciences NCHRP-IDEA program, a conventional sign detection algorithm based on color, shape, and texture has been developed to process the images of signs. The developed algorithm has been enhanced by digital image processing and deep-learning methods, such as convolutional neural networks (CNN), generative adversarial networks (GAN), and region-based convolutional neural networks (RCNN). In this paper, a method is designed to process the images obtained using unmanned aerial vehicles (UAV), employing a model UAV, to further develop of our current research. The preliminary test shows that it is promising to use a UAV and machine learning to develop an expedited infrastructure condition evaluation following natural disasters because of its automatic and non-contact nature. The preliminary outcomes show the detection rates have satisfying FN rates and very low FP rates. Besides, the designed algorithm and data-gathering method provides a real-time and on-site computation capability that reduces the quantity of data to be stored by filtering out unnecessary data instantly.