For a park with a small budget, the choice has been between paying for expensive thermal drones and having rangers stare at night-time video for hours. A student’s prototype tries to remove that trade-off. It costs about $300 and uses software to tell people from elephants.
The student is Anika Puri, who was 17 and from Chappaqua, New York, when she won a $25,000 scholarship from the Davidson Fellows programme for the project, called ElSa (Elephant-Savior). Indian outlets are carrying the story again this week. The award itself was announced on 25 August 2022, and Puri has since graduated from the Massachusetts Institute of Technology.
The idea began in India. On family trips she was struck by rows of ivory jewellery and statues in Mumbai’s markets, despite the 1990 global ban. She then looked at how poachers are found. Rangers fly drones with heat-sensitive cameras, and operators must watch the live video continuously, a tedious and error-prone task. Software that spots shapes also struggles, because animals and people look tiny from high altitudes.
Puri’s approach looks at movement rather than shape. She noticed that humans and elephants differ in speed, turning patterns and other motions. Her model classifies a figure using speed, group size, turning radius and number of turns. On a thermal video dataset called BIRDSAI, collected in four African national parks, it detected humans with 90.9% accuracy, which she says is four times better than the best existing methods. The Davidson Institute’s press release puts the gain at nearly 70 percent, so the two figures are measured differently.
The hardware is cheap. ElSa uses a $250 FLIR ONE Pro thermal camera with 206×156 pixel resolution that plugs into an off-the-shelf iPhone 6. Her entry puts the cost of a high-resolution thermal camera at $4,800. FLIR’s own write-up says drone-based AI systems already in use can cost as much as $10,000 per drone.
The results come from recorded footage, not from rangers using the system in a live park. Independent comment has been limited. Jasper Eikelboom, an ecologist at Wageningen University in the Netherlands, told Smithsonian Magazine: “It’s quite remarkable that a high school student has been able to do something like this.”
At the 2022 Regeneron International Science and Engineering Fair, Puri received the $10,000 Peggy Scripps Award for Science Communication. The Davidson programme named only 21 winners across the country that year. To train her classifier, she used 516 time series extracted from videos of humans or elephants in motion.
