Learning-Deep-Learning

- Deep Radar Detector

May 2019

tl;dr: Use the range-doppler-channel data cube to perform 4D detection task (range, doppler, azimuth and elevation).

Overall impression

This is a splendid paper with great overall introduction to conventional radar detection algorithms and how they replaced conventional blocks with CNN. This work also used radar calibration data with phase shift augmentation to train NN.

With enough data and richer annotation, this work could be extended to detect multiple objects, and maybe even regress the size of the object, if the resolution is sufficiently high.

This is further enhanced by Qualcomm’s deep radar perception which directly regresses a bbox from the range-doppler-azimuth tensor.

Background

Key ideas

Technical details

Notes