Machine-learning radars may be coming to automotive

May 25, 2016 // By Peter Clarke
The IMEC research institute (Heverlee, Belgium) plans to make future sensors, and specifically radar sensors, devices that extract useful information locally and even become learning machines.

IMEC is already working with automotive radar market leader Infineon Technologies AG at 79GHz in 28nm CMOS (see Infineon, IMEC collaborate on 79GHz CMOS radar ). Now it wants to go to a yet smaller wavelength and add machine learning to the back end of its sensors said Wim van Thillo, program director for perceptive systems at IMEC, speaking at the IMEC Technology Forum.

Van Thillo said his group is already working on a 140GHz chip. At this frequency the wavelength is 2.2mm and his group is aiming for more than 4GHz of bandwidth from a chip measuring 1 square millimeter, he added.

The advantages will include higher distance and angular resolution at lower power in a much smaller system size with the radar able to include the antenna-on-chip. In addition to angle and distance the radar is able to provide speed information via a mini-doppler effect. The use of multiple antennas integrated on to the chip will result in enhanced Doppler resolution and a better depth resolution.

The signal processing that will be needed to extract speed information is likely to be taken further with the use of algorithms for pattern recognition and automatic learning. As a result Van Thillo invisages a time wen the radar will be able to recognize and distinguish the signature of pedestrians, bicycles and cars from their mini-Doppler signatures.

Massimiliano Maranella, program manager for millimeter-wave perceptive systems, said that 28nm CMOS is being used again and that in simulation at least the signals look good enough to use. However, it acknowledged that a 28nm CMOS platform would not be suitable for long-range (300m) automotive radar.

A complication is that the plastics used in bumpers and metallic paints tend to absorb the energy at 140GHz, Maranella said.

However, while the frequency may be used in shorter-range gesture recognizing systems the learning can be extended across a broader set of perceptive systems including automotive.

For example,