A behaviour-centric digital twin framework for animal monitoring: from managed enclosures to landscapes

For decades, the way we watch animals has been a game of fragments. In zoos, wildlife reserves, and rehabilitation centers, researchers rely on fixed cameras to observe behavior. A camera in the corner sees the left side of a pond; another on a pole sees the right. Each lens captures a slice of the world, but no single camera sees the whole picture. When scientists try to understand how an animal uses its space, they often have to stitch these disjointed views together by hand, or worse, they mistake a blind spot for a place the animal simply does not like. This approach is not just incomplete; it is often disruptive. Installing new cameras in an occupied enclosure requires drilling, cabling, and repeated human presence, which stresses the very animals being studied. Ideally, the monitoring system should be perfect before a single tool touches the ground, and the data should reflect the animal's actual world, not just the camera's limited view.

This is where the concept of a "digital twin" enters the story. Imagine a perfect, virtual copy of a physical place, built with the same dimensions, textures, and lighting as the real thing. In engineering, these twins are used to test machines before they are built. In ecology, they are beginning to help scientists understand populations, but they have rarely been used to watch individual animals in their daily habitats. A new framework developed by researchers at Vantara, a wildlife rescue and rehabilitation center in India, aims to fill this gap. They have created a system that turns a zoo enclosure into a living, interactive map. This system does not just record video; it understands the three-dimensional space, knows exactly what each camera can and cannot see, and translates every movement into a precise location on a map that represents the animal's true world.

The researchers applied this framework to two specific enclosures at their center: one for American black bears and another for pygmy hippopotamuses. For the black bear paddock, which covers nearly 1,000 square meters, they built a detailed virtual replica using satellite images, drone photos, and ground surveys. This digital model includes the terrain, a pond, rock walls, climbing structures, and even the specific trees and logs that might block a view. They placed virtual cameras in the exact spots where the real cameras were installed, mimicking their height and angle. Before the system could be trusted with real data, the team ran a rigorous self-test. They generated thousands of synthetic, computer-generated movements for two virtual bears, creating a "ground truth" of where the animals were supposed to be. They then fed these fake movements through their detection system to see if the software could accurately recover the positions.

The results of this simulation were precise. When the system looked at the data from a single camera, it could pinpoint a bear's location with a median error of about 19 centimeters. However, the system's true power emerged when it combined the views from multiple cameras. By fusing the data from different angles, the error dropped to just 7 centimeters. This fusion allowed the system to see 94.8% of the time the animals were active, effectively eliminating the "blind spots" that plague traditional monitoring. Crucially, the system could also distinguish between a space that was truly empty and a space that was simply hidden from view. In older methods, if a camera could not see a corner of the enclosure, the software might assume the animal avoided that corner. This new framework marks those hidden areas as "unobservable," ensuring that the data reflects what was actually seen, not what was missed.

Beyond just tracking movement, the framework allows for a level of planning that was previously impossible. The researchers used the digital twin to analyze the camera network before any hardware was installed, a process that would have been impossible in a real, occupied enclosure without disturbing the animals. They calculated that the five-camera setup they eventually installed would make 85.1% of the paddock visible to at least one camera, with 57.4% visible to two or more cameras simultaneously. The system identified that one specific camera, positioned near the den door, was a specialist that saw very little of the general area but was essential for watching the entrance. It also revealed that another camera, located at the pool end, was the most critical for overall coverage; if that one failed, the visible area would drop significantly. Furthermore, the twin could predict when the sun would shine directly into the lenses, causing glare that washes out the image. It showed that the south-facing cameras would be blinded by the sun for up to ten hours a day during the winter solstice, a problem that could now be solved by adjusting the mounting angle or using infrared lighting before the cameras were even turned on.

The ultimate goal of this work is to create a continuous story for each animal, from rehabilitation to release. Currently, an animal's life is often studied in two disconnected phases: the time spent in a managed enclosure and the time spent in the wild. This framework proposes a bridge between the two. The same digital architecture used to monitor a bear in a zoo could be adapted to monitor that same bear after it is released into a vast landscape, using GPS collars and camera traps instead of fixed CCTV. This would allow scientists to maintain a single, consistent record of an individual's behavior, comparing how it used space in captivity with how it navigates the wild. While the current study focused on the precision of the enclosure monitoring, the design is built to scale up to these larger, wilder environments.

The researchers emphasize that their findings are based on a highly accurate simulation and a retrospective analysis of an already installed camera network. The system has not yet been tested on live, real-time footage of the bears in this specific study, though the underlying detection software is already running on the center's network. The success of the simulation proves that the mathematical and geometric foundations are sound, establishing a margin of error that is small enough to trust for behavioral analysis. By moving the design and testing phase into a virtual world, the framework ensures that the physical installation is correct the first time, sparing the animals the stress of repeated construction. It transforms animal monitoring from a collection of fragmented, two-dimensional video clips into a unified, three-dimensional measurement of life, where every step is accounted for, and every blind spot is known.