AI startup builds a Wi-Fi motion sensor that knows when you’ve fallen over
China’s Perspicace hired stunt performers to fall over to train sensor
One evening a few years ago, Ken Yip’s grandmother fell off her bed in a Shanghai flat. Frail and injured, the elderly woman spent the whole night lying on the floor, unable to get up and ask for help.
It was an unfortunate episode -- but Yip, the CEO and founder of Shanghai-based startup Perspicace, thinks it could have been avoided. At the Microsoft AI Summit in Hong Kong, he showed me what he calls a Wi-Fi Bio-detector: a device that looks like a standard wall socket combined with a switch box.
Yip says the device combines a Wi-Fi router with AI algorithms that are designed to detect various kinds of human movement. Here’s how it works: In a room without people, a Wi-Fi router distributes signals in a stable and tidy manner. But when someone enters the room, the signals become disturbed -- creating noise that is instantly registered on the detector.
So for example, if you’re standing, you disturb the signals in a certain kind of way. But if you then fall, your movement creates a different kind of disturbance entirely. Yip’s challenge was to find a way to translate the data he was gathering into something useful -- working out whether this signal disturbance he detected meant a person was standing, sitting, even breathing heavily.
Yip, who returned to China after a career in AI with the French gaming giant Ubisoft, started working on the idea in 2013. Yip says he hired a large number of people to work in teams of three -- collecting data 24-7. One person would go about his daily life, while the others record his movement. While that happens, a machine in the background would register any changes in Wi-Fi signals.
But that’s not all. To perfect their understanding of each motion, Yip said he recruited stunt performers. With the falling motion, for instance, he asked them to act out different types of falls -- ranging from tripping to fainting. Their machine learning algorithms need to learn from some 3,000 data sets for each type of movement. Together, it took five years for Perspicace to accumulate the over 5 million data sets it has today.