Computer Vision for Wildlife Monitoring and Conservation
Background
Effective wildlife conservation depends on understanding how animal populations change over time. Conservation practitioners need reliable information about population size, distribution, movement patterns, and long-term trends in order to assess the status of species and make informed management decisions. Collecting this information, however, is often challenging. Many wildlife species inhabit remote areas, occur at low densities, and often move across vast landscapes, making direct observation difficult and expensive.
Over the last two decades, camera traps have transformed wildlife monitoring by providing a non-invasive method for observing animals in their natural habitats. Networks of cameras can operate continuously for weeks at a time, generating detailed records of wildlife activity across large geographic areas. As camera trap deployments have become more widespread, they have also created a new challenge: the sheer volume of imagery produced can be overwhelming to process manually.
Our work focuses on developing computer vision systems tailored to wildlife monitoring in Kenya. By combining animal detection, species classification, and individual animal re-identification, we aim to build tools that help conservationists process camera trap data more efficiently, generate reliable population insights, and support evidence-based conservation decisions.
Accomplishments
In partnership with Mugie Conservancy in Laikipia, Northern Kenya we are developing an application to process large sets of camera trap images. The application takes a large set of camera trap images, runs the processing behind the scenes using a cascade of models and generates visualizations. It also provides an interface to download summary statistics such as species counts over time.
The application is currently in its later stages of development and will soon be deployed for field use to help conservationists in their monitoring efforts.
Next Steps
For our next steps, we will be aiming to do the following:
- Adapting Individual Re-Identification to the Camera Trap Setting
- Deploying DSAIL-Annotate Wildlife Analysis Application
- Leveraging Synthetic Data To Improve Re-Identification Performance
Re-identification is the task of identifying whether two images belong to the same individual. Currently, the developed methods for re-identification of Grevy's and Plains Zebras are feasible only for bespoke census datasets with images collected by humans using DSLR cameras during census exercises. Our aim is to adapt the re-identification methods to camera trap images that suffer from various quality issues such as motion blur, occlusions, low resolution, difficult image framing and variable lighting conditions.
We are continuing to develop and refine wildlife data processing tools for use by conservation practitioners. The goal is to move these systems from research prototypes toward practical deployment, where they can support real-world wildlife monitoring workflows and be iteratively improved through field use.
A major bottleneck in re-identification research is the limited availability of labelled individual-level data. To address this, we are exploring the use of synthetic data generated from 3D animal representations to augment existing datasets. By simulating variations in pose, viewpoint, and lighting, synthetic data offers a scalable way to improve model robustness for individual re-identification tasks.