中国科学家为自动驾驶汽车和国防领域开发出世界上最快的视觉芯片 || Chinese scientists create world’s fastest vision chip for autonomous cars, defence

Chinese scientists say they have developed the world’s first brain-inspired vision chip with two pathways that could have broad uses in autonomous vehicles and defence.

The team from Tsinghua University in Beijing say their chip, Tianmouc, has a record-breaking image processing speed. They said it had achieved high-speed sensing of up to 10,000 frames per second, a dynamic range of 130 decibels, and excellent spatial resolution.

Data flow is reduced by 90 per cent compared with traditional high-speed imaging chips and power consumption is low.

“This is a perception chip, not a computational one, based on our original technical route,” project leader Shi Luping, a professor with the university’s Centre for Brain-Inspired Computing Research, said in a statement on Tsinghua’s website on Thursday.

“Firstly, it balances speed and dynamic performance in vision chips and introduces a novel computational method that diverges from existing machine vision strategies,” Shi said.

“Secondly, this approach mimics the human visual system’s dual pathway, enabling decision-making without complete clarity,” he said.

“The chip provides new directions for advancements in autonomous driving and defence sectors. It is poised to tackle some of the most challenging issues faced today, paving the way for numerous novel applications.”

The team reported on the development in the journal Nature on Thursday, with Shi and Professor Zhao Rong the corresponding authors of the study.

The visual perception terminals in intelligent systems process vast volumes of data and must be able to handle extreme events like sudden driving hazards, significant light changes at tunnel entrances, and intense nighttime flash interference.

Traditional visual chips frequently fail or experience delays in these extreme scenarios due to limitations in power consumption and bandwidth.

The scientists said Tianmouc was inspired by the human visual system. It analyses visual information along two pathways: one for cognition and another for rapid response.

Explaining the difference between the chips, the scientists gave the example of flying objects, which humans – and Tianmouc – will subconsciously dodge even if they do not see them clearly. Traditional machine vision chips, however, often need to see clearly before making decisions.

Of all the information collected by the vision sensor, Tianmouc uses colour, absolute intensity, precision and resolution in a cognition-oriented pathway for accurate recognition.

Spatial difference, temporal difference and speed are meanwhile used in an action-oriented pathway for a fast response.

The team said it developed software and algorithms based on the chip, which were deployed on a Chevrolet for road testing, confirming its capabilities in various extreme scenarios.

A statement on the Tsinghua website said the chip had shown low latency and high performance in real-time perceptual reasoning across a range of challenging environments, showing significant potential for applications in unmanned systems.

Dr Wang Taoyi, a member of the research team, told state broadcaster CCTV on Friday that Tianmouc had achieved an 0.1 millisecond delay in autonomous driving. “This is just 1/300th of the 30ms delay experienced with traditional cameras,” Wang said.

The vision chip represents significant progress from another brain-inspired chip developed by the Tsinghua centre. A computing chip, Tianjic handled multiple AI algorithms in an energy-efficient way to control a riderless bicycle, executing tasks such as turning, accelerating and avoidance based on voice commands. The team reported their findings in Nature in 2019.

via News - South China Morning Post (author: Zhang Tong)
 
 
Back to Top