Radio detection and ranging (a.k.a radar) sensors are commonly used in various applications due to their reliability in adverse lighting and weather conditions, long detection range, and ability to detect an object's relative velocity. While various techniques and waveforms can be used to perform radar sensing, frequency-modulated continuous wave (FMCW) radars are very common for low-cost sensing applications. The latest generation of FMCW radars, mm-Wave radar operates at frequencies ranging from 77-81GHz with bandwidths of up to 4 GHz; such high bandwidths and frequencies allow these sensors to have 20 x better range resolution (down to 4cm) and 3x better velocity resolution.

A common FMCW radar signal processing pipeline is featured above. From the figure, the 3 key steps of FMCW radar signal processing can be summarized as follows:

1. Chirps at Transmitter (Tx) and Receiver (Rx): For each radar frame, a series of chirps are transmitted. These chirps reflect off of objects in the environment, and the reflected chirps are received by the radar.

2. Dechirping and IF Signal Generation: The transmitted and received chirps are then fed through a mixer to determine a the intermediate frequency (a.k.a IF). This is the frequency difference between the transmitted and received chirp, and can easily be used to measure the range of different objects. 

3. Range-Azimuth Response: Using an operation known as a 2D fast Fourrier transform (a.ka. 2D-FFT), we compute the Range-Azimuth response which shows the received signal power a various ranges and azimuth (horizontal) angles.

We utilize a simplified U-net architecture in order to convert low resolution radar range-azimuth responses into higher resolution "lidar-like" 2D point clouds. The high level parameters of our model are described below:

• Input: 40 x 64 x 48 normalized tensor corresponding to a normalized range-azimuth responses (in polar coordinates) from 40 chirps in a radar frame. For each chirp, we record 64 adc samples across 4 receivers. The azimuth FFT is zero padded to 64 elements, but we only keep the parts corresponding to +/- 50 degrees, resulting in 48 azimuth bin for each range

• Output: 64 x 48 segmentation mask derived from the lidar ground truth data

• Architecture: U-net

• Loss Function: 0.9 weighted binary cross entropy (BCE) and 0.1 DICE loss