Home → Environment → Animals
A new AI system tracks 46 facial points to spot signs of fatigue in working dogs.
Dogs can’t tell us when they’re exhausted. But scientists are teaching artificial intelligence to recognize when they’ve had enough, simply by looking at their faces.
Researchers in Israel have developed a system that analyzes ordinary smartphone videos, tracking 46 tiny points around a dog’s eyes, ears, nose, and mouth. By studying how these facial features change, the AI learned to distinguish working dogs filmed before strenuous activity from those filmed afterward.
For now, the main application isn’t figuring out when our best friends still want to run or need a rest. Rather, it could help handlers decide when search-and-rescue or explosives-detection dogs need a break.
Dogs experience the world largely through smell. Their noses can separate complex mixtures of odors and detect traces that humans would never notice. That’s why dogs are so good at detecting things like explosives, drugs, or even diseases like cancer or malaria.
When a detection dog works, it repeatedly draws air through its nose, sampling the environment for chemical clues. Sniffing is an active process, and maintaining that rhythm matters.
Unfortunately, dogs also depend heavily on panting to regulate their body temperature. Unlike humans, who cool themselves primarily by sweating, dogs release much of their excess heat through evaporation from their respiratory passages.
So, sniffing and panting end up competing for the same time.
A landmark 2003 experiment by Israeli researchers demonstrated the consequences. Dogs searching for explosives after strenuous exercise panted more, sniffed less frequently, and performed worse at detecting their targets. In a 2025 study, researchers tested 11 explosives-detection dogs, comparing their search performance with and without prior exercise. Without exercise, the dogs completed searches in an average of around 30 seconds. After exercise, that rose to 45 seconds.
Stay ahead with ZME Science and subscribe.
Please check your inbox and confirm your subscription.
Heat adds another complication. A 2024 study found that working dogs needed about 3.5 times the concentration of one explosive odor to detect it under extremely hot, humid conditions compared with milder conditions.
This can be very problematic for work dogs. They have to work hard, sometimes in hot or dangerous environments, but continued exertion could undermine both performance and welfare.
This is where the new tech comes in.
The University of Haifa team, led by researchers including Anna Zamansky, studied four professionally trained working dogs: two search-and-rescue dogs and two detection dogs.
Handlers filmed the animals using ordinary smartphones. Some recordings captured dogs that had rested for at least 24 hours. Others showed them immediately after 20 to 45 minutes of intensive searching or detection work. Overall, they had 177 clips.
Rather than simply looking for an open mouth or a drooping expression, the system followed how key facial points moved and changed relative to one another. Overall, the software identified 46 anatomical landmarks on each dog’s face, including points around the eyes, ears, nose, and lips.
The team then tested several machine-learning methods, including neural networks that analyze sequences of facial movements.
The AI learned to distinguish dogs filmed before and after strenuous work. For dogs it had already encountered, it correctly classified about 78% of recordings. For entirely unfamiliar dogs, accuracy dropped to 69%.
The results suggest that strenuous activity leaves subtle, measurable traces on a dog’s face, even if those changes aren’t obvious to the human eye. The researchers haven’t yet established whether the system can measure fatigue itself, but the findings point toward a future in which a simple camera could help reveal when a dog needs a break.
Search-and-rescue dogs, police dogs, and other working animals often operate in environments where stopping for a physical examination is impractical. A quick camera-based assessment could eventually enable dog owners to figure out whether an animal needs rest.
But interpreting a dog’s face is complicated.
Panting can reflect heat, exercise, stress, or excitementa. Ear position and facial anatomy differ dramatically between breeds. Even a familiar expression may mean different things depending on the situation.
Indeed, another 2026 study involving researchers from the same laboratory found that people spent substantially more visual attention interpreting the faces of short-muzzled dogs than those with more typical muzzle shapes.
Still, the central idea is compelling, and there’s definitely room for improvement. In fact, there’s a much larger scientific effort underway to use AI to understand what animals are experiencing and communicating.
The same University of Haifa team has trained AI to recognize pain in cats, using subtle changes around their eyes, ears, and mouths. They’ve also explored how a detection dog’s tail movements might reveal when it has found a scent.
Elsewhere, researchers with Project CETI are using machine learning to decipher the complex clicking patterns of sperm whales, while the Earth Species Project is analyzing animal vocalizations, from zebra finches to killer whales.
We’re still far from a universal animal translator. But increasingly, technology is offering us new ways to recognize what animals are expressing, especially when the signals are too subtle for our senses to notice.
The study was published in Nature Scientific Reports.
Dr. Andrei Mihai is a geophysicist and founder of ZME Science. He has a Ph.D. in geophysics and archaeology and has completed courses from prestigious universities (with programs ranging from climate and astronomy to chemistry and geology). He is passionate about making research more accessible to everyone and communicating news and features to a broad audience.
© 2007-2025 ZME Science – Not exactly rocket science. All Rights Reserved.
© 2007-2025 ZME Science – Not exactly rocket science. All Rights Reserved.

