Sydney company The Ripper Group and the University of Technology Sydney have built an AI system, SharkSpotter, that identifies sharks in live drone footage with about 90% accuracy, compared with under 20% for traditional aerial patrols, according to UTS.
Shark surveillance at Australian beaches has long depended on people scanning the water from helicopters, light planes or towers, a method limited by glare, fatigue and the speed of the aircraft. Pairing drones with a trained image recognition model offers a cheaper and more consistent way to watch the surf, and gives lifesavers a tool that can also respond to an emergency.
How SharkSpotter works
SharkSpotter runs on the Westpac Little Ripper drones that patrol beaches in Australia. The drone streams live video to an operator while a deep learning model analyses the footage frame by frame.
The model has been trained to separate sharks from 16 other categories of objects in the water, including dolphins, rays, surfers, swimmers and boats. When it detects a shark, it alerts the operator, who can then warn lifeguards and people in the water.
The accuracy gap
UTS puts SharkSpotter’s detection accuracy at about 90%. Human spotters in aircraft fare much worse. The Conversation cited detection rates of 17.1% for helicopters and 12.5% for fixed wing aircraft.
Unlike a human observer, the model does not tire over a long patrol, and the drone can hover over a suspicious shape rather than flying past it.
The Lennox Head rescue
The drones carry more than a camera. In January 2018, a Little Ripper drone dropped a flotation pod to two young swimmers stuck in a rip off Lennox Head on the NSW north coast. The rescue took 70 seconds from launch. UTS estimated that a rescue launched from the beach would have taken six to seven minutes.
According to The Conversation, the drones can also carry an electronic shark repellent alongside the flotation pod.
The partnership
SharkSpotter combines The Ripper Group’s drone fleet with machine learning research from UTS. The project is one of the better known Australian examples of applied AI in public safety, a field that has drawn far less attention than chatbots and consumer software.
Jack’s take
Most AI news I see is about chatbots and valuations. This is a drone, a model and a flotation pod reaching two kids in a rip in about a minute. To me, that is what AI was supposed to be for.