Vehicle Detection With Automotive Radar Using Deep Learning on Range-Azimuth-Doppler Tensors

November 2019

tl;dr: Deep radar perception on radar FFT data.

Overall impression

Directly regress oriented bbox (maybe it is even upright bbox) on front radar. This is similar to deep radar detector. This work is succeed by Qualcomm’s NIPS 2019 paper on camera and radar fusion.

Limitation: The dataset only contains highway driving, and only have upright bbox.

The paper reiterated that when a transformation is mathematically known, it is easier to code it into the neural network.

Key ideas

Technical details