Researchers have developed a low-power artificial intelligence system that can identify drones while using dramatically less electricity than conventional computer-based systems.
The technology could make it easier to continuously monitor unauthorized drones around military sites, factories and other sensitive locations.
The system, called UAV-NAS, was developed by Doyeon Kim, an undergraduate researcher in the Department of Electronic and Electrical Engineering at Sungkyunkwan University in South Korea.
The research was published in IEEE Transactions on Industrial Informatics.
Drones are becoming increasingly common around the world. They are used for photography, deliveries, agriculture, inspections and many other purposes.
But their rapid growth has also created security concerns, particularly when unauthorized drones enter restricted airspace.
Artificial intelligence can help security systems recognize drones and determine what they are doing. However, powerful AI models usually require significant computing resources. Running these models on standard computer processors can consume a lot of electricity.
That creates a problem for small surveillance systems operating outdoors or in remote locations. If a drone-monitoring device relies on batteries, a power-hungry AI system can greatly reduce how long it can operate before it needs to be recharged.
Kim developed UAV-NAS to address this problem. Instead of having engineers manually design the structure of the AI model, the system uses a technique that allows computers to search for an efficient AI architecture automatically.
In simple terms, the computer looks for a version of the AI that can perform the necessary drone-identification tasks without carrying around unnecessary computational complexity.
The researchers then adapted the resulting AI model to run on an FPGA, or field-programmable gate array. An FPGA is a type of semiconductor chip that can be configured for particular computing tasks. Because the hardware can be optimized for a specific job, it can sometimes perform calculations much more efficiently than a general-purpose computer processor.
Testing showed that the system could accurately identify different types of drones as well as their flight status. At the same time, it reduced power consumption by 88.7% compared with running the AI system on a conventional CPU.
The large reduction in electricity use could make the technology particularly valuable for surveillance systems that need to operate continuously. Low-power drone detectors could potentially be installed in outdoor locations, industrial facilities, military environments and other places where access to a constant power supply may be limited.
The achievement is particularly notable because Kim carried out the research while still an undergraduate student. He has been conducting research on FPGA semiconductor technology in a university laboratory since his junior year and was also selected for an internship at Rebellions, a South Korean AI semiconductor company.
Kim is expected to graduate with his bachelor’s degree in August. In September, he will begin a fully funded Ph.D. program at Purdue University in the United States, where he plans to continue studying semiconductors and artificial intelligence.
The research demonstrates how designing AI software and computer hardware together could make intelligent systems much more energy-efficient, potentially allowing sophisticated AI to operate in smaller devices far from powerful data centers.
