Houston, Jul 17 (The Conversation) Artificial Intelligence in search and rescue may not surpass human accuracy, but it significantly outpaces humans in speed. Recent advancements in using computer vision and machine learning to analyze drone imagery have shown promise in rapidly assessing building and road damage following hurricanes or tracking wildfire progressions. These developments highlight AI's potential in locating missing persons after floods.
Machine learning systems excel by scanning high-resolution drone images in under a second, compared to the one to three minutes required by human analysts. Additionally, drones can capture vast amounts of imagery, beyond what humans can feasibly review in the crucial initial hours of a search when survivors may still have a chance.
However, current AI systems have limitations. As researchers specializing in disaster robotics, our experiences in searching for flood victims and other situations have shown inadequacies in current AI implementations.
Despite this, AI has a role to play when combined with human efforts while searching for flood victims. The key lies in collaboration.
The Potential of AI Searching for flood victims is a unique challenge within wilderness search and rescue. Machine learning aims to rank images by the likelihood of containing victim indicators and direct search-and-rescue teams to focus on specific areas within those images. If responders identify signs of a victim, they can provide GPS locations to field teams for further investigation.
This image ranking is managed by a classifier, an algorithm trained to recognize similar objects—like cats, cars, or trees—from training data to identify them in new images. In a rescue context, a classifier might detect human activity signs such as litter or gear for wilderness search-and-rescue teams, and potentially pinpoint the missing person.
With drones able to produce vast quantities of imagery, a classifier becomes essential. A 20-minute drone flight can yield over 800 high-resolution images, and ten such flights, considered minimal, produce over 8,000 images.
Reviewing each image for only 10 seconds would demand over 22 hours of effort. Even when distributed among multiple “squinters,” humans can overlook image sections and experience cognitive fatigue.
An ideal AI system would scan images in full, highlight those with the highest probability of finding victims, and mark image areas for responders to check. It could also assess if the area warrants special attention from search-and-rescue teams.
Challenges with AI Implementation Although AI seems an ideal fit for this task, contemporary systems suffer high error rates. Programming them to err on the side of caution can lead to an overabundance of false candidate locations.
This could inundate squinters or complicate work for search-and-rescue teams, navigating debris and challenging terrain to verify each location.
Three primary obstacles complicate developing AI for finding flood victims. First, current computer vision systems can identify visible people in aerial imagery, but flood victims often appear differently than hikers or fugitives, being obscured, camouflaged, or submerged. This visual complexity heightens error risks for classifiers.
Second, machine learning needs extensive training data, but datasets featuring humans trapped in debris or mud are scarce, increasing classification discrepancies.
Third, drone images often capture oblique views instead of vertical shots. This misalignment means that the GPS coordinates in images may not align with the drone's location. Calculating precise GPS coordinates is possible if the drone’s altitude and camera angle are known, but these data are commonly unavailable. Resulting inaccuracies demand extra search time from teams.
Enhancing AI and Human Collaboration Despite the challenges, AI combined with human judgment allows search-and-rescue teams to efficiently prioritize and focus their efforts.
In flood scenarios, human remains might be entwined with vegetative debris. AI can identify large debris clusters likely to hide remains. A strategic approach involves noting GPS locations of flotsam buildup, as victims could be located within these aggregations.
An AI classifier can also detect debris indicative of remains, such as unnatural colors and construction debris with straight lines and right angles.
While responders typically survey banks and floodplains on foot, classifiers can help prioritize locations during the critical initial hours and days, and subsequently confirm no areas were overlooked as teams traverse the difficult terrain. (The Conversation) SCY SCY
(Only the headline of this report may have been reworked by Editorji; the rest of the content is auto-generated from a syndicated feed.)
