Hadar heats up race for better night-time computer vision, AV performance
Machine learning helps take fuzz out of thermal imaging
A machine learning-assisted approach to thermal imaging could transform the night vision of autonomous vehicles and other nascent robots.
The quest for driverless cars has hit some roadblocks, including efforts to get guidance systems to see in the darkness at night. Although thermal imaging has been around for decades, a scattering of thermal radiation can create fuzzy images, limiting their use in automated navigation.
To crack the problem, a team of researchers have reached for machine learning and a stack of material data.
The team, led by Purdue University professor of electrical and computer engineering Zubin Jacob, developed "heat-assisted detection and ranging," which is being abbreviated to HADAR.
To try to make sense of fuzzy signals received by thermal detectors, the team developed a custom library of data describing the emissivity of every material they were likely to encounter. Emissivity is a measure of the material's effectiveness at emitting thermal radiation relative to a perfect black body, an idealized physical body that absorbs all incident electromagnetic radiation. The goal was to disentangle signals coming directly from an object and reflected interference signals, while also helping identify the material.
The researchers trained a neural-network model with data from this material library, helping it to process heat signals from the infrared camera in a bid to figure out the temperature, emissivity and texture of each object in the field. It also learned from its own data to map the world through heat signals, according to a paper published in Nature this week.
"We develop HADAR estimation theory and address its photonic shot-noise limits depicting information-theoretic bounds to HADAR-based AI performance. HADAR ranging at night beats thermal ranging and shows an accuracy comparable with RGB stereovision in daylight," the paper said.
- SpaceX grounds Falcon 9 carrying first private lunar lander
- Pyro-brainiacs set new record with waste-heat-into-electricity study
- US army boffins use AI to spot faces in the dark
- In current affairs news: Teen boffin with lots of potential crafts electric honeycombs out of oil
In an accompanying article, Manish Bhattarai, a researcher scientist at the Los Alamos National Laboratory, said: "Although HADAR is a nascent technology, its potential applications are many and varied. It seems clear that the system will find immediate applications in the autonomous driving and robotics sectors. But it could also be applied in national security and emergency-response settings, in which the success of a mission can hinge on the responder's ability to navigate under conditions of near-zero visibility."
Bhattarai also said the technology could help with body-scanning at airports, wildlife monitoring and in geoscience research.
"The scalability and passive nature of HADAR will no doubt inspire future imaging and vision technologies. However, the system is not without its challenges. The greatest barriers lie in the cost of the equipment, and in hardware-level issues, including the fact that the system must be calibrated on the fly," he said. ®