The ranking is done by a classifier, which is an algorithm that learns to identify similar instances of objects – cats, cars, trees – from training data in order to recognize those objects in new images.
For example, in a search-and-rescue context, a classifier would spot instances of human activity such as garbage or backpacks to pass to wilderness search-and-rescue teams, or even identify the missing person themselves.
A classifier is needed because of the sheer volume of imagery that drones can produce. For example, a single 20-minute flight can produce over 800 high-resolution images. If there are 10 flights – a small number – there would be over 8,000 images. If a responder spends only 10 seconds looking at each image, it would take over 22 hours of effort. Even if the task is divided among a group of “squinters,” humans tend to miss areas of images and show cognitive fatigue.
The ideal solution is an AI system that scans the entire image, prioritizes images that have the strongest signs of victims, and highlights the area of the image for a responder to inspect. It could also decide whether the location should be flagged for special attention by search-and-rescue crews.
Where AI falls short
While this seems to be a perfect opportunity for computer vision and machine learning, modern systems have a high error rate. If the system is programmed to overestimate the number of candidate locations in hopes of not missing any victims, it will likely produce too many false candidates. That would mean overloading squinters or, worse, the search-and-rescue teams, which would have to navigate through debris and muck to check the candidate locations.
Developing computer vision and machine learning systems for finding flood victims is difficult for three reasons.
One is that while existing computer vision systems are certainly capable of identifying people visible in aerial imagery, the visual indicators of a flood victim are often very different compared with those for a lost hiker or fugitive. Flood victims are often obscured, camouflaged, entangled in debris or submerged in water. These visual challenges increase the possibility that existing classifiers will miss victims.
Second, machine learning requires training data, but there are no datasets of aerial imagery where humans are tangled in debris, covered in mud and not in normal postures. This lack also increases the possibility of errors in classification.
Third, many of the drone images often captured by searchers are oblique views, rather than looking straight down. This means the GPS location of a candidate area is not the same as the GPS location of the drone. It is possible to compute the GPS location if the drone’s altitude and camera angle are known, but unfortunately those attributes rarely are. The imprecise GPS location means teams have to spend extra time searching.