Chinese scientists have helped create a new safety system for automated driving systems that can sometimes react to hazards more quickly than the human brain.
The slower reaction time of machines compared with the human brain has been a long-standing safety concern.
For example, an automated vehicle travelling at 80km/h (50mph) will take half a second to respond to a hazard in front of it, compared with the 0.15 seconds the human brain needs to react – meaning the vehicle will have travelled another 13 metres (43 feet) before stopping.
Even advanced processors have been far slower than humans at analysing a high-definition image to discern what is moving and where it is going.
This delay has created a fundamental safety concern about robots, drones and autonomous vehicles.
However, researchers from Britain, mainland China, Hong Kong, Saudi Arabia and the United States said they had now developed a faster reaction system.
The research was published online in the peer-reviewed journal Nature Communications.
Co-corresponding author Gao Shuo, an associate professor in the School of Instrumentation and Optoelectronic Engineering at Beihang University, said: “Our approach demonstrates a 400% speed-up, surpassing human-level performance while maintaining or improving accuracy through temporal priors.
“We do not completely overthrow the existing camera system; instead, by using hardware plug-ins, we enable existing computer vision algorithms to run four times faster than before, which holds greater practical value for engineering applications.”
The team tackled the problem by mimicking human vision, which reacts to moving objects when it encounters sudden obstacles rather than trying to process the whole visual field.
At the heart of the research was hardware that followed the “filter-then-process” dynamic with a two-dimensional synaptic transistor array – essentially a highly sensitive “motion detection” chip.
This transistor possessed three key capabilities: the ability to detect image changes in just 100 microseconds (far faster than human perception); to retain motion information for over 10,000 seconds; and operate for more than 8,000 operational cycles without its performance worsening.
Once it captured a frame, it was able to ignore the full image and only register the key changes and flag all “moving objects”, which were then passed to standard vision algorithms for detailed analysis.
This process was more than 10 times faster than conventional approaches, the paper said.
The team tested the system in different situations, including autonomous driving, drone navigation and the use of robotic arms.
They said the hardware had processed motion data four times faster than current state-of-the-art algorithms and, under ideal lab conditions, even surpassed human performance.
They reported a 213.5% improvement in the ability to detect hazards when driving and a 740.9% boost in robotic arms’ ability to grasp objects.
In real-world tests, the system’s efficiency decreased slightly but was still better than previous autonomous driving systems, and the researchers said the roughly 0.2-second improvement could reduce braking distance by 4.4 metres in a vehicle travelling at 80 km/h.
Gao said: “In a traffic accident, these 4 metres often determine whether a collision occurs or it’s just a close call.”
For small drones, the system helped cut reaction time by at least one third, greatly enhancing endurance and performance.
“This project will undoubtedly advance in-depth collaboration with Chinese automotive and drone companies,” Gao said.
“We hope to equip autonomous vehicles with this ‘hardware-level reflex’ system, enabling them to respond more sensitively than humans when handling sudden road conditions, thereby fundamentally enhancing the safety of unmanned systems.” – South China Morning Post
