US 11,885,907 B2
Deep neural network for detecting obstacle instances using radar sensors in autonomous machine applications
Alexander Popov, Kirkland, WA (US); Nikolai Smolyanskiy, Seattle, WA (US); Ryan Oldja, Redmond, WA (US); Shane Murray, San Jose, CA (US); Tilman Wekel, Sunnyvale, CA (US); David Nister, Bellevue, WA (US); Joachim Pehserl, Lynnwood, WA (US); Ruchi Bhargava, Redmond, WA (US); and Sangmin Oh, San Jose, CA (US)
Assigned to NVIDIA Corporation, Santa Clara, CA (US)
Filed by NVIDIA Corporation, Santa Clara, CA (US)
Filed on Mar. 31, 2020, as Appl. No. 16/836,583.
Claims priority of provisional application 62/938,852, filed on Nov. 21, 2019.
Prior Publication US 2021/0156960 A1, May 27, 2021
Int. Cl. G01S 7/295 (2006.01); G06T 7/246 (2017.01); G06T 7/73 (2017.01); G01S 7/41 (2006.01); G01S 13/931 (2020.01); G06N 3/08 (2023.01); G06V 10/764 (2022.01); G06V 10/82 (2022.01); G06V 20/58 (2022.01); G06V 20/64 (2022.01)
CPC G01S 7/2955 (2013.01) [G01S 7/414 (2013.01); G01S 7/417 (2013.01); G01S 13/931 (2013.01); G06N 3/08 (2013.01); G06T 7/246 (2017.01); G06T 7/73 (2017.01); G06V 10/764 (2022.01); G06V 10/82 (2022.01); G06V 20/58 (2022.01); G06V 20/64 (2022.01); G06T 2207/10044 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30261 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method comprising:
based at least on data representative of a projection image of RADAR data representative of sensor detections generated using one or more RADAR sensors of an ego-actor, computing, (i) using a class confidence head that is of a neural network and comprises a plurality of classification channels, a first output representative of a confidence map for each of the classification channels representing likelihood that at least one individual pixel of one or more pixels belongs to a corresponding class of a plurality of classes of one or more detected objects, and (ii) using an instance regression head that is of the neural network and comprises a plurality of regression channels for each class of the plurality of classes, a second output representative of, for at least one individual regression channel of the plurality of regression channels for a first class of the plurality classes, one or more regressed values of one or more object instances of the one or more detected objects that belong to the first class and are represented by the at least one individual pixel; and
generating, based at least on the first output and the second output of the neural network, at least one bounding shape corresponding to at least one detected object of the one or more detected objects.