An aerial picture can show where a herd is. A three-dimensional movement record can show how its members turn, separate and regroup. Researchers working at Ol Pejeta Conservancy in Kenya are combining drones, computer vision and geometric reconstruction to make those richer observations possible.[1][2]
European research projects have been applying artificial intelligence, autonomous drones and photogrammetry to these questions. In October 2026, Fondazione Bruno Kessler reported results from WildDrone, whose technologies were developed and field-tested with partners at Ol Pejeta Conservancy in Kenya. The systems use imagery to reconstruct individual animals in three-dimensional space and estimate their position, orientation, size and changes in location over time. Related researchers at the University of Bristol have also released a substantial aerial wildlife-tracking dataset from Ol Pejeta.[1][2]
The task is to turn sightings into measurements that can be compared across time. Positions, orientation and trajectories can give conservation researchers a more detailed account of behaviour than an isolated image or a simple count.
Why two dimensions can mislead
What the Kenya evidence contains
| Record | Grounded interpretation |
|---|---|
| 22 WildLive aerial sequences | A set of annotated video recordingsSource 2 |
| 19,139 frames | Individual images within those sequencesSource 2 |
| 215,800 boxes and masks | Frame-level annotations, not separate animalsSource 2 |
| 291 tracklets | Observed identity segments, not 291 confirmed unique animalsSource 2 |
| WildDrone field development | Real-world 3D monitoring tests at Ol PejetaSource 1 |
A bounding box drawn around an animal in a photograph is an imperfect representation of life on the ground. An elephant closer to the drone may occupy more pixels than a larger animal farther away. A giraffe's apparent length shifts dramatically as it turns. Trees can hide individuals, and two animals can temporarily overlap in the image. A straightforward count of boxes may therefore become a poor account of bodies and movements.[1][2]
Photogrammetry reconstructs spatial information from imagery using geometrical relationships, including how subjects appear from different vantage points. Machine-learning systems can help identify animal-shaped regions and match their appearances across time. Together, they offer a route from an isolated photograph to a temporally connected representation of a moving group.
The field work adds complications that tidy laboratory demonstrations avoid. Animals do not stay on marked paths. Light changes, bushes obscure views, battery life runs down, and a drone cannot occupy every viewpoint at once. A system that tracks twenty visible animals well may still lose an individual that disappears behind the others and emerges later. Researchers therefore need both datasets with known annotations and careful field checks.[1][2]
A tracking dataset with 19,139 frames
The University of Bristol's WildLive dataset contains 22 UAV-acquired 4K video sequences from the Ol Pejeta site, totalling 19,139 frames. Its annotations include 215,800 bounding boxes and segmentation masks covering 291 tracklets of zebras, giraffes and elephants, along with 84 sparse point tracks. A tracklet is a continuous observed segment of one object's journey, not necessarily a distinct animal across a whole day.[2]
The WildLive research and WildDrone project are complementary evidence streams rather than interchangeable performance claims. The former makes tracking examples available to researchers; the latter reports tested 3D monitoring approaches and their field-development context. Combining them conceptually helps explain the frontier, but their metrics should not be silently pooled as if they came from one identical trial.[1][2]
From research flights to repeatable surveys
Monitoring methods also affect the behaviour they measure. Drone noise, shadows and flight paths may disturb different species in different ways. Testing that response alongside tracking quality would show when aerial observation offers an advantage over ground-based surveys.[1][2]
Fondazione Bruno Kessler says experience from WildDrone is feeding into WildBotics, a further programme involving multiple sensing platforms and trials in the Trentino Dolomites. This is a roadmap for continued development rather than proof that all proposed autonomy has already been deployed. Each new terrain brings different canopy, GPS, terrain and weather challenges.[1]
The practical gain would be continuity: finding the same animals again after a turn, overlap or brief disappearance, then placing their movements in physical space. Better continuity makes a flight more than a collection of pictures; it becomes a record researchers can interrogate.[1][2]
Sources & further reading
- Robots, AI, and 3D reconstructions to help conserve endangered animal species Fondazione Bruno Kessler · 2 Oct 2026 · Accessed 26 Oct 2026
- WildLive dataset University of Bristol · 28 Jan 2026 · Accessed 26 Oct 2026