Researchers develop machine learning method to accurately identify objects using microwaves for use in self-driving cars

 According to foreign media reports, engineers from

The new method is based on machine learning and eliminates the middleman, skipping the step of creating images for manual analysis and directly analyzing pure data. In addition, the method can also determine the optimal hardware configuration, revealing the most important data while discovering what the most important data is. In a proof-of-concept study, the setup took dozens of measurements instead of the hundreds or thousands usually required to correctly identify a set of 3D numbers.

In the study, the researchers used a metamaterial antenna that can shape microwave wavefronts into many different shapes. In this case, the metamaterial is an 8×8 square grid, each of which contains electronic structures that can be dynamically adjusted to block or transmit microwaves.

In the new object recognition technology, a radio wave source (back panel) creates a wavefront (center panel) that is made of a metamaterial screen that allows the radio waves to pass through in some places but not others (front panel). Machine learning then finds the waveform that illuminates the most useful features of the object, an approach that improves recognition accuracy while reducing computing time and power consumption.

In each measurement, the smart sensor selects multiple square grids through which microwaves can pass, creating a unique microwave pattern that reflects off the identified object and back to another similar metamaterial antenna. The sensing antenna also uses an active square grid to add more options to shape the reflected waves. A computer then analyzes the incoming signal and attempts to identify the object.

By repeating this process thousands of times for different situations, the machine learning algorithm eventually discovers what information is most important and what settings on the transmitting and receiving antennas are best for collecting that information.

Once trained, the machine learning algorithm focuses on a small set of settings that help it distinguish “useful” (wheat) data from useless data (chaff), reducing the number of measurements, time, and computing power required. Traditional microwave imaging systems typically require hundreds or even thousands of measurements, while the current system only needs fewer than 10 measurements to “see” an object.

Whether this improvement can expand the technology to more complex sensing applications remains to be seen, but researchers have already tried to use the new concept to optimize hand movement and gesture recognition for next-generation computer interfaces. There are many other areas that need to improve microwave sensing technology, and the small size, low cost and easy manufacturing characteristics of this metamaterial make it useful in future devices. (All pictures in the article are from Duke University)