Algorithms for species conservation

Monitoring is necessary to track the status of wildlife populations, to assess whether they are stable or under pressure. Naturalists observe zebras, giraffes, jaguars, and other animals over the course of months and years using drones or camera traps. The problem is that the animals are constantly on the move, disappearing and reappearing elsewhere. To be able to capture this highly dynamic environment, animals must be individually identifiable. Specifically, this means that it must be possible to clearly distinguish a single jaguar from other jaguars, and a single giraffe from other giraffes. While conventional technologies do provide accurate results in this regard, they are too slow and require too much computing power.

The Flight Robotics Group led by tenure-track professor Aamir Ahmad, based at the Institute of Flight Mechanics and Controls (IFR) at the University of Stuttgart, together with collaborators, among them the Eötvös Loránd University, Budapest and the Max Planck Institute for Intelligent Systems, created an algorithm that makes the re-identification of wild animals easier, faster, and more reliable.  The scientists explain how it works in the journal "Methods in Ecology and Evolution".

DOI:  https://doi.org/10.1111/2041-210x.70332

Like a fingerprint: Every pattern is unique

The acronym RAPID stands for “Real-Time Animal Pattern Re-Identification on Edge Devices.”  RAPID uses a 100 percent reliable characteristic for recognition: the coat patterns of wild animals. After all, whether it's the spots on a giraffe or the stripes on a zebra—the pattern is always unique, just like a human fingerprint.

But how exactly does RAPID work? “First, we need a reference database,” says Aamir Ahmad. This database contains images of wild animals – such as zebras – whose individual identities are already known. Over time, newly observed individuals can also be added to this database, allowing it to grow during monitoring. Once deployed in the field, for example on image data from a research drone circling over the savanna, RAPID scans these images for distinctive visual cues in each zebra’s stripe pattern. “To do this, we use descriptor vectors, a kind of mathematical profile,” explains András Zábó, researcher at the Eötvös Loránd University, Budapest and first author of the publication. He developed RAPID during his research stay in the Flight Robotics Group. The algorithm compares the descriptor vectors of the drone-observed animal with the descriptor vectors of known animals stored in the database. “In this way, we can identify a newly observed individual — provided it is represented in the reference database — in a fraction of a second,” says Zábó.

A promising module: fast, precise, and practical

The research team tested RAPID on six datasets: four public datasets containing images of, among other animals, Amur tigers, and two new datasets containing videos of zebras and jaguars. The footage of the zebras was captured by drones flying over the savanna in Mpala, Kenya; the jaguar footage comes from the Jocotoco Foundation and consists of camera trap videos recorded in the rainforest in Ecuador. On the four public datasets, RAPID achieved accuracies ranging from 89 to 99 percent; on the new datasets, accuracy was 80 percent (zebras) and 93 percent (jaguars). The algorithm also demonstrated its strengths in terms of speed. On a standard PC, it processed 40 to 60 cropped query images per second, and on a basic edge device, about ten images per second. “That's an important point: Our recognition system works even on hardware with very limited processing power and without a GPU,” says Ahmad.

Open-source and modular: A key component for wildlife monitoring and ecological analyses

The AI tool is available as open source and has a modular design. If park rangers or research groups want to use RAPID, they can easily integrate it into their own drones, camera traps, airships, or other monitoring devices. The only requirement: the animals being observed must have patterned coat; for example, the technology does not work with elephants. The new algorithm is an important technical component for future wildlife monitoring and ecological analyses. “With RAPID, it's much easier for us to determine whether a particular individual is repeatedly sighted in a specific area. We can also observe whether its behavior changes over time as its environment changes, or whether an injured animal continues to move and interact normally,” Ahmad explains. Next, he plans to further develop the AI so that it can also recognize other wild animal species without depending on their coat patterns. In addition, the algorithm is to be made even more robust so that it functions reliably even under particularly difficult conditions like partially occluded individuals. It is also conceivable that, with the help of RAPID, entirely new databases will eventually be created.

About the research project „Wildcap“:

The new RAPID algorithm was developed as part of the “Wildcap” research project (duration: May 2021 – April 2026). The Flight Robotics Group at the Institute for Flight Mechanics and Controls (IFR) at the University of Stuttgart collaborated on the project with researchers in Kenya and Hungary. Partners included, among others, Princeton University in the United States and Hortobágy National Park in Hungary. The goal was to use artificial intelligence and autonomous aerial robots—a drone and an airship—to monitor endangered wildlife species.  “Wildcap” was funded by the Cyber Valley Research Fund.