South African NPO Shark Spotters, the Institute for Communities and Wildlife in Southern Africa at UCT and a Swiss-based machine learning company will teach a computer to spot sharks to help prevent attacks on some of the country's most dangerous beaches.
It's an ambitious idea: building a low-cost automated shark spotting system that will pick up what a fatigued human eye could miss, from a fixed live-feed camera.
It means an unlikely partnership between Shark Spotters and PatternLab, a Swiss company which specialises in data science and computer-aided vision.
But where the Swiss team may lack experience in tracking sharks, they more than make up in tech.
Dr. Krzysztof Kryszczuk, a senior pattern recognition specialist from PatternLab has a wealth of knowledge in teaching machines to track objects and behaviours.
From a crowded room in the Save Our Seas Shark Education Centre in Kalk Bay, the kite-surfing tanned Kryszczuk gave a presentation showing a year’s worth of research that laid the groundwork for the project.
“Sharks don't show up every day. So how do we build data? We need images and videos of sharks,” said Kryszczuk.
Fixed cameras, located at a Shark Spotter lookout point, recorded the tests and began to teach the machine to pick up the appearance model of the 2-meter sized Joshua, which looks like a small blurry dot on screen.
We zoomed in for you.
Over time it will learn to differentiate between the pattern of a real shark versus something like a wad of kelp that drifts in with the tide.
They bring with them 15 years’ worth of knowledge and experience on shark behaviour with more than 2,500 successful shark-human preventions.
Their experience has proven to be vital to training the machine as they have the know-how to spot sharks under different weather conditions. A bright sunny day might seem ideal for people to spot sharks, but on a camera its a different story competing with glare and luminosity for example. The ocean can also look completely different when its cloudy, or when the wind is blowing, which the machine will have to learn from scratch.
“These were factors we didn’t know about, but to the spotters these factors were intuitively known,” said Kryszczuk.
Machine learning and fixed cameras are a much more cost-effective way of managing shark spotting than other forms of technology, said Sarah Waries CEO of Shark Spotters. They thought drones could work, but in order to do the same work as a machine, you would need to have a fleet of drones that can't fly when its windy.
Using this technology could also help Shark Spotters at other beaches. The current system is dependent on spotting stations being located high up on cliffs, so spotters can see into the water. A camera on a pole, could be a cheap easy solution, if it works out.
“We want it to be able to pick up a number of different marine animals. But this will require ‘training’ to be able to differentiate between sharks, whales, dolphins etc,” said Waries.
The research project will run for 2 years. The project is funded by the Eurostars programme, an international scheme that supports innovative projects led by research and development. In SA its funding is administered by the department of science and technology, and in Switzerland by Innosuisse.
According to the conditions of the funding, the project needs to be commercialised after three years.
Also from Business Insider South Africa: