Imec adds machine learning to its smart radar system
The ultra-fine resolution of the MIMO setup will also allow the detection of micro-skin movements related to vital signs to enhance applications such as non-contact driver monitoring or patient monitoring.
imec’s 140GHz radar-on-chip prototype system is small and offers high radar performance, especially in resolution and motion sensitivity. The radar operates up to 10m range, with 15mm range resolution and 10GHz of RF bandwidth. Multiple antenna paths are incorporated to enable a complete (virtual) 1×4 MIMO configuration for angular target separation. The transceiver chip features on-chip antennas, and are integrated in 28nm bulk CMOS technology, ensuring a low-cost solution at high volume production. The machine learning capabilities are intended to demonstrate the feasibility of the radar to detect and classify small motions based on Doppler information.
Being insensitive to lighting conditions and preserving privacy (a radar can so far not recognize humans), a radar solution has particular advantages over other types of motion sensors. Imec developed a specific machine learning algorithm based on a multi-layer neural network including an LSTM layer and using supervised learning to train the inference model by using in-house labelled recordings of more than 25 people, including several captures for each of 7 different gestures. Against the experimental dataset, the model classifies the recorded 7 gestures and predicts the right gesture at least 94% of the time. Vital signs can also be measured with very high precision due to the high radio frequency.
imec is also currently building a 4×4 MIMO radar system, for which a new generation of radar chips is under development – incorporating the TX and RX as separate chips.