What level of vision does an autonomous car need to be truly safe on public roads? If we’re going to place our lives in the hands of machines and software, then can we agree that the enabling visual hardware must exceed human 20/20 vision and even achieve “superhuman perception” by pairing high-resolution cameras with a flood of sensor-driven data, perhaps collected with LiDAR?
Waymo robotaxis, for instance, crash less often than human drivers, according to recent data, and the company is rolling out its service with test runs in a growing number of US markets. But there’s a distressing blind spot – literally – with all this tech that is worrisome at best, dangerous at worst.
A Vision Problem That Needs Addressing
Volvo. EX90. LiDARVolvo
The key phrase there is test runs, because human safety drivers (or “autonomous specialists,” as Waymo calls them) need to be behind the wheel, ready to take over in the event of a glitch or if superhuman perception is compromised. That’s good, because CarBuzz discovered some research that suggests vision systems for autonomous cars still have a ways to go before they are ready for prime-time applications, safely transporting thousands or millions of passengers a year, without a human at the wheel.
The latest study comes from two South Korean college professors who found that vision systems onboard autonomous vehicles often struggle to accurately read speed-limit signs after years of exposure to the elements. Researchers Seong Tae Kim from Kyung Hee University and Hong Joo Lee from Seoul National University of Science and Technology developed a framework to simulate normal wear and tear on traffic signs “to show how it can mislead AI recognition systems.”
How Hard Is It To Read Damaged Road Signs?
The team trained a generative image-to-image translation model to learn and then recreate the visual characteristics of damaged and undamaged traffic signs. The result was artificially generated signs that looked naturally degraded. The researchers then printed clean and damaged speed-limit signs and photographed them at different distances, viewing angles, and indoor and outdoor conditions. “The resulting images remained effective at misleading the traffic-sign classifier” used in autonomous vehicles.
Through this testing, the team also discovered that vision systems could be trained to read aged or damaged road signs, “suggesting that simulated natural deterioration could be used to identify and strengthen weaknesses in vision systems.” Kim said building reliable artificial intelligence requires more than improving average performance.
“It requires continuously identifying where AI systems fail, understanding why those failures occur, and using those insights to make the systems more robust.”
–Seong Tae Kim, Associate Professor At Kyung Hee University
Protecting Against Bad Actors Physically Altering Signs
Highway speed limit signAn Errant Knight
The other study from earlier this year comes from Johns Hopkins and the University of California, Santa Cruz, in which researchers tested how vision systems used by autonomous vehicles and drones respond when written language on signs is deliberately manipulated.
Sure enough, researchers found that altered signs can mislead self-driving cars and drones.
In simulated driving scenarios, a modified road sign, for instance, caused a self-driving car to interpret the altered text as a directive and proceed despite pedestrians being present, according to a story in TechRadar. If you’re having visions of Wile E. Coyote slamming into a wall because Road Runner painted a tunnel on it, you’re not alone.
Zeekr RTWaymo
CarBuzz Insight – Why This Matters:
Autonomous cars are not yet intelligent enough to accurately process signs that have been damaged or intentionally altered, so there’s still work to do. Important research is underway that could make next-generation vision systems more robust, to discern correct road signs from those that have been damaged or physically changed.
If natural aging of signs becomes a problem for these vision systems, local governments hoping to ensure safe roads will have to pony up for replacement signs more frequently, or be on high alert for bad actors spray-painting signs. There’s a lot at stake. As the Korean researchers say, “misinterpreting a traffic sign could potentially lead to unsafe decisions.”
For all that goes into modern AI systems, it still boils down to a computer system running processes that are easily thrown off by simple things humans take for granted. This study highlights the fundamental issue that will always plague such systems – difficulty adapting to the unknown.
Source: SeoulTech
