Drones are already used in search and rescue operations. But planning a search route is more of an art than a science. Artificial intelligence can change this.
If a tourist gets lost in Scotland’s rugged highlands, rescue teams sometimes fly a drone to look for signs of their route. Trampled grass, lost clothing, a food wrapper. In this case, it is critical to correctly determine the area to search, given the vast terrain and limited operating time of the device.
Testing a Machine Learning System
Traditionally, experienced operators use intuition and statistical “search theory” to do this. It is a strategy that was used to detect German submarines during World War II. Jan-Hendrik Evers and his team from the University of Glasgow decided to test whether a machine learning system could work more efficiently.
Evers grew up in Scotland. There he enjoyed skiing and hiking. So he clearly understands the challenges associated with rescue operations in these places.
“As a child, I had nothing to do except spend time outdoors or sit in front of the computer,” he says. “I ended up doing a lot of both.”
To start, Evers took a set of data related to search and rescue operations around the world. Here you could find the age of the missing person, whether he was hunting, riding or hiking, whether he suffered from dementia, and where he was eventually found – near water, in a building, in an open area, under a tree or on a road.
The researcher trained an AI model on the dataset and loaded Scotland’s geodata into it. The model runs millions of simulations to identify the routes a missing person would most likely take in given specific circumstances. The result is a probability distribution. A kind of heat map—that indicates priority search areas.
Deep Learning to Design Routes
Using this map, the team demonstrated that deep learning could help design more efficient routes for search drones. In a study published on arXiv, the team tested their algorithm on two common search patterns:
“lawnmower” – the drone combs the target area, moving in stripes,
and an algorithm similar to the Evers algorithm, but less advanced in terms of working with probability distribution maps.
In virtual testing, Evers’ algorithm outperformed both approaches in two key metrics. The distance a drone would have to fly before finding a missing person, and the likelihood that the person would be found.
Combining the area and the existing algorithmic approach helped find the person in 8% and 12% of cases, respectively. The approach proposed by Evers showed an efficiency of 19%. If the system proves successful in real-world conditions, it could speed up response times and save more lives in situations where every minute counts.
Experts believe that deep learning will help develop more efficient routes and find missing people faster in the wild, depending on how suitable the environment is for drone searches (for example, exploring dense forested areas is more difficult than bushland).
Risks
But don’t forget about the nuances. The success of such a planning algorithm will depend on the accuracy of the probability maps. And if you rely too much on them, you risk drone operators spending too much time exploring the wrong areas.
Evers says the next important step is to get as much training data as possible. To do this, he hopes to use GPS data collected from later rescue operations for modeling. Essentially, this will allow the model to understand the relationship between the location where the person was last seen and the location where they were eventually found.
However, transaction records are not always detailed enough to be workable. In the broader spectrum of AI applications, platforms like Woo Live Casino demonstrate how AI can revolutionize various fields, including live gaming.
UAVs in Expansion
Unmanned aerial vehicles are becoming increasingly common in the search and rescue world. But it’s still a relatively new technology, and the rules governing its use are still changing.
In the USA, for example, UAVs must always be in the operator’s field of view. Meanwhile in Scotland, operators are prohibited from being more than 500m away from a drone. These rules are intended to avoid accidents such as a drone crashing and putting people in danger, but in emergency situations such rules severely limit the ability of ground rescuers to search for evidence.
“Often the problem is not technical, but regulatory,” says Kovar. “Drones can do much more than what we are allowed to do with them.”
Evers hopes models like his can further expand the capabilities of drones. He is currently in talks with Police Scotland’s Air Support Unit to see what it will take to test and deploy his system in real-world conditions.
