Observing group-living animals with drones and computer vision
-
Loading... - Kanoro Okeyo
- 23 Mar 2023
- 71 Views
- 0 Like
- 0 Comment
A drone is flying over a herd of plains zebras in central Kenya. It is flying quite high so that the animals are not bothered by it. These zebras are really interesting for collective and spatial behaviour studies, as the researchers Ben Koger and Blair Costelloe, who are monitoring the drone, say. The plains zebras live in multi-level societies: small groups of females and a male combine to form larger herds of dozens of animals. This social and spatial structure could influence behavioural processes such as decision-making and information sharing and have implications for understanding our own complex societies. Traditionally, it has been very difficult to conduct this kind of research. But new techniques their team has developed using imaging drones and artificial intelligence open up new possibilities.
To explore animal groups such as zebras or gelada monkeys, Ben Koger, Blair Costelloe, Iain Couzin, and other researchers from the Max Planck Institute of Animal Behavior, the "Centre for the Advanced Study of Collective Behaviour" (CASCB) at the University of Konstanz, and Aarhus University developed a new method for collecting data about animal behaviour and the animals' surrounding natural physical landscape using drones and computer vision.
The researchers use imaging drones to record entire groups of animals in natural settings. Behavioural ecologist Blair Costelloe describes the method: "We created an analytical pipeline that lets us take aerial drone footage and extract information about the locations, movement, and behaviour of the animals. We can measure their spatial distribution and their behavioural states and get rich information about their surroundings, including the 3D-structure of the environment."
Bringing tracking from the lab to the field
Previously researchers mostly got high precision data sets about animal group dynamics in highly-controlled labs conditions where you could repeat experiments over and over. But the team asked themselves: "Could we use imaging drones and new computer algorithms to take the same lab approaches but bring them into the natural landscapes?"
It is possible -- but several challenges had to be solved: "We were often recording 20 or more different individuals at a time. Quantifying where each of the individuals is in a single half hour video observation as a human would take weeks," Ben Koger explains. "The first challenge was how could we automatically detect the animals we were interested in?" The solution was training powerful deep learning algorithms. The second challenge: The researchers were interested in the animals' movements, and yet the videos they recorded included not only animal movement but also drone movement and distortions from the hilly landscape they were filming over. All those different elements needed to be untangled before they could get meaningful data. Read More…
