Imaging technologies play a vital role in the emerging autonomous vehicle ecosystem. There is a wide consensus among different players in this industry that a combination of several long-range (over 100m) depth sensing modalities is imperative for the viability of self-driving cars. Optical technologies allowing accurate depth imaging at such ranges are practically restricted to LIDAR (laser rangefinder) sensors such as Velodyne, which has become the golden standard in research platforms. However, LIDARs suffer from several critical shortcomings, in part due to their reliance on a rotating mechanical assembly: they are prohibitively expensive for the consumer market, have insufficient temporal and spatial resolution, do not provide an accurate velocity reading, and operate at a shorter range in adverse weather conditions. Radio-frequency (RF) sensors complement the optical modality in autonomous vehicles. Specifically, millimeter wave multiple-input multiple-output (MIMO) radar is a mature technology providing accurate range and velocity images at relatively long distances. MIMO radar relies on a phased array of transmitting antennas and digital receivers, contains no moving parts, and can penetrate much denser fog and rain compared to the optical counterparts. The current weakness of this technology is that in order to achieve sufficient spatial resolution, multiple receiver channels are required, dictating the high cost of the device. The ability to maintain high resolution and quality images using a smaller number of receiver channels has the potential to significantly reduce the cost of this technology and increase the commercial viability of automotive digital MIMO radars. We have developed a proof-of-concept system allowing major optimization of mm-wave MIMO radar systems. Our deep learning-based methodology allows to reduce the number of receive channels, the number of frequency points, and the total transmission time with negligible impact on the reconstructed image quality.
