A Hong Kong facial-recognition firm whose artificial intelligence (AI) cameras tackled prison suicides has won a $4m Smart Traffic Fund grant to develop congestion prediction technology.
The project, one of four funded under the latest round of the Smart Traffic Fund, will use AI and pan-tilt-zoom cameras to help Transport Department (TD) stay ahead of traffic jams.
Traffic is a huge potential application for AI in Hong Kong says Ivy Li, founder of Wildfaces, a Hong Kong firm with “smart city” applications in 75 cities around the world.
“There’s too many cars, the roads are too narrow,” she says.
Right now, Transport Department uses human eyes to check the traffic situation in the city, a time consuming process.
The Wildfaces project will be able to predict congestion in real time, adding new layers of information to Hong Kong’s camera network without upgrading the camera infrastructure.
“So if there is illegal parking near a tunnel entrance during rush hour, there will be congestion there, so we can alert [Transport Department] ahead. Or if a truck is trying to turn around, taking too many turns, or there’s too many pedestrians crossing the road longer than the green light, or a broken traffic light, we can tell them in advance.”
Li says the system will have considerable advantages over services like Google’s traffic map, which, she says, relies on people logging in or sharing location info with Google and “is usually only half correct.”
“By the time you see it, you’re already stuck in the traffic. And it cannot tell you the cause of the congestion,” she says.
“On the move” cameras
Li says Hong Kong’s traffic-cam network is in fact quite limited compared to, for example, Dubai, mainland China or other Asian cities.
Under the 12-month project, Wildfaces will be able to extend the city’s camera view without adding a single lens, using AI together with pan-tilt-zoom camera functions for the first time in Hong Kong.
“There’s really not many cameras to play with, so we like to extend, to make use of existing cameras to expand their view, using ‘pan-tilt-zoom’ to expand the view, look further, look wider with the same camera infrastructure,” she says.
So while, for example, Wildfaces’ Cyberport neighbour Logistics and Supply Chain Management (LSCM) has developed fixed camera systems for the police, nobody else has yet deployed AI on moving cameras. “We expect to file at least two patents from this,” says Li. “Nobody else can manage ‘on the move’.”
Li says 20 years’ experience with visual AI gives her a more pragmatic approach than many AI or “deep learning” start-ups on the scene today. She says AI projects may be over-engineered towards unnecessary learning.
“Our technology approach is to be able to mimic complex human intelligence. Adults don’t just keep learning and learning, you use past experience, use some logic to solve the problem,” she says.
For example, in the prison project, Li’s team preferred to make a human study of risky behaviours before programming the AI, rather than employing a lengthy machine learning phase.
Correctional Services Department has not commented on how many suicides or fights that project prevented – and Li is not privy to such information. “They won’t tell us that,” she says, but says the project is being discussed for a broader rollout.
“I’ve been developing visual AI for the past 20 years so we know how to make use of AI to automate these time-consuming human processes,” she says.
For traffic, Wildfaces takes a similar approach, using its own experience and logic to skip length learning stages.
“So for traffic, if we need to know which type of car, or is it a red taxi or a green taxi, or a truck or what sort of truck, we might need to learn a little, but we really try to minimise that work.”
Privacy has been an issue in Hong Kong since before protestors chopped down smart lampposts with chainsaws, and Li says she is well aware of the challenges.
“TD likes our work because we can do low resolution, the cameras are capable of high resolution but we try to reduce the lowest possible resolution to still understand the situation,” she says.
“So if someone is trespassing [on the highway or tunnel], it’s not our intent to see who the person is, just see there is someone walking on the highway and alert the relevant authorities”.
The firm blurs out any personal information such as faces or licence plates, and doesn’t store video. “We are not surveillance,” she says. “We only look at the incidents.”
While TD won’t commit to long-term use of the system, Li says there are plenty of customers in facility management and construction. “They need the same thing, they have roads to manage, small roads entering the facility or estate. In our [Smart Traffic Fund] proposal, we told them, ‘great if TD can come in but if they don’t then fine, we will be able to survive!'”