Machine Learning for Wildlife Conservation with UAVs

Poaching has increased in the last decade, particularly poaching of elephants and rhinoceroses. Current levels of poaching of elephants and rhinoceroses will lead to their extinction in the next 10 years. Unmanned aerial vehicles (UAVs) can be flown to spot poachers or to locate animals.

We are working to automatically detect poachers and animals in thermal infrared images using deep learning techniques. We are also interested in planning UAV patrol routes. We continue to be interested in recruiting students to join us in this project.

SPOT is a tool that automatically detects poachers in long wave thermal infrared UAV videos. Tests of SPOT have been run by Air Shepherd at a testing site in South Africa, where training exercises take place. This video is sped up 20 times and shows a 30 minute test at the site using Microsoft Azure on a slow internet connection

We are working to automatically detect poachers and animals in thermal infrared images using deep learning techniques. We are also interested in planning UAV patrol routes. We continue to be interested in recruiting students to join us in this project.

SPOT is a tool that automatically detects poachers in long wave thermal infrared UAV videos. Tests of SPOT have been run by Air Shepherd at a testing site in South Africa, where training exercises take place. This video is sped up 20 times and shows a 30 minute test at the site using Microsoft Azure on a slow internet connection.

A drone is following poachers through the AirSim-W Africa environment. The normal drone view is on the left, and the corresponding thermal infrared simulation is on the right. SPOT is looking for poachers in the simulated thermal infrared image, and poacher detections are shown in blue boxes. Now, we can see how SPOT would work in the field using a simulated environment.

The Benchmarking IR Dataset for Surveillance with Aerial Intelligence (BIRDSAI, pronounced "bird's-eye") is the long-wave thermal infrared dataset containing nighttime images of animals and humans in Southern Africa used for this research. The dataset allows for benchmarking of algorithms for automatic detection and tracking of humans and animals with both real and synthetic videos. 

More information can be found on the dataset homepage. The data can be downloaded from the Labeled Information Library of Alexandria. We also provide a brief video description. 

Where should the limited number of conservation drones fly to best protect animals? Can anything be done to deter poaching if park rangers are unable to reach a target quickly? We address these questions using game theory, and show that strategic signaling can deter poaching both theoretically and in simulation.