Houston: In search and rescue, AI is less accurate than humans, but much faster.
Recent success in applying computer vision and machine learning to drone images to quickly determine damage to buildings and roads after a hurricane suggests that artificial intelligence is valuable in searching for missing people after a flood.
Machine learning systems usually take less than a second, scan high-resolution images from drones, and for humans it takes 1-3 minutes. Furthermore, drones often produce more images than is humanly possible in the critical first time of searches where survivors may still be alive.
Unfortunately, today's AI systems have not reached the task.
We are robot retreats studying drone use in disasters. Our experiences and many other events looking for flood victims indicate a lack of current AI implementation.
However, this technology can play a role in finding flood victims. The important thing is AI-Human Collaboration.
The possibilities of AI
Finding flood victims is a type of search and rescue in the wilderness that presents its own challenges. The goal of machine learning scientists is to rank which images have signs of casualties and show where search and rescue personnel should focus on those images. If the responder sees a victim's indication, they pass the location of the GPS in the image and search for teams in the field to check.
Ranking is done by a classifier. This is an algorithm that learns to identify similar instances of an object. – Cat, car, tree – From training data to recognize these objects in new images.
For example, in the context of search and rescue, classifiers find instances of human activity, such as garbage or backpacks, passing through wilderness search and rescue teams, or identifying the missing person themselves.
Because of the enormous amount of images that drones can generate, classifiers are required. For example, one 20-minute flight can produce over 800 high-resolution images. If there are 10 flights – few – There are over 8,000 images.
If the responder only uses 10 seconds to examine each image, it will take more than 22 hours of effort. Even if the task is divided into groups of “squids”, humans tend to miss areas of the image and show cognitive fatigue.
The ideal solution is an AI system that scans the entire image, prioritizes the image with the strongest signs of the victim, and emphasizes the area of the image for the responder to inspect. You can also decide whether to flag a location for special attention by search and rescue crew.
There's a lack of AI
This seems like a great opportunity for computer vision and machine learning, but modern systems have high error rates. If the system is programmed to overestimate the number of candidate sites in the hopes of not missing the victim, it can generate too many false candidates.
It means a squid overload, or even worse, a search and rescue team that needs to lure debris and mud to see where the candidate is located.
Development of computer vision and machine learning systems to find flood victims is difficult for three reasons.
One is that while existing computer vision systems can certainly identify people who see in aerial images, visual indicators for flood victims are often very different compared to visual indicators for lost hikers and fugitives. Flood victims are often obscure, camouflage, entangled with debris, or submerged in water. These visual challenges increase the likelihood that existing classifiers will miss victims.
Second, machine learning requires training data, but there is no dataset of aerial images where humans are tangled in debris and covered in mud, rather than normal posture. This lack also increases the likelihood of errors in classification.
Third, many of the drone images that searchers often capture are diagonal views rather than looking straight down. This means that the GPS location of the candidate area is not the same as the GPS location of the drone. It is possible to calculate the GPS position if you know the altitude of the drone and the angle of the camera, but unfortunately these attributes are rare. Inaccurate GPS locations means that teams need to spend extra time searching.
How AI can help
Fortunately, humans and AI work together, search rescue teams can use existing systems to narrow down and prioritize images for further inspection.
In the case of a flood, human remains may be entangled with vegetation and debris. Thus, the system can identify chunks of debris large enough to include the body. A common search strategy is to identify the GPS locations of the locations Flotsam collected, as the victims could be part of these same deposits.
AI classifiers can find debris commonly associated with debris, such as artificial colors and construction debris with straight lines and 90 degree angles.
The responders spot these signs as they walked systematically through the riverbanks and flood plains, but the classifier helped them prioritize areas in the first hours and days where there could be survivors, and later confirmed that they didn't miss an area of interest when the team navigated the difficult landscape on foot. (conversation)
