US 11,710,254 B2
Neural network object detection
Shubham Shrivastava, Sunnyvale, CA (US); Punarjay Chakravarty, Campbell, CA (US); and Gaurav Pandey, College Station, TX (US)
Assigned to Ford Global Technologies, LLC, Dearborn, MI (US)
Filed by Ford Global Technologies, LLC, Dearborn, MI (US)
Filed on Apr. 7, 2021, as Appl. No. 17/224,181.
Prior Publication US 2022/0335647 A1, Oct. 20, 2022
Int. Cl. G06T 7/73 (2017.01); G06N 3/08 (2023.01); G06V 10/82 (2022.01); H04N 23/90 (2023.01)
CPC G06T 7/74 (2017.01) [G06N 3/08 (2013.01); G06V 10/82 (2022.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30252 (2013.01); H04N 23/90 (2023.01)] 20 Claims
OG exemplary drawing
 
1. A system, comprising:
a first image sensor positioned to obtain a first image of an object;
a second image sensor positioned to obtain a second image of the object;
a computer including a processor and a memory, the memory storing instructions executable by the processor to:
input the first image to a neural network that outputs a first six degree-of-freedom (DoF) pose of the object from a perspective of the first image sensor;
input the second image to the neural network that outputs a second six DoF pose of the object from a perspective of the second image sensor;
determine a pose offset between the first and second six DoF poses by determining a difference between respective three-dimensional (3D) bounding boxes for the object determined based on the first and second six DoF poses;
determine a first projection offset by determining a difference between a two-dimensional (2D) ground truth bounding box for the object and a first 2D bounding box generated from the first six DoF pose;
determine a second projection offset by determining a difference between the 2D ground truth bounding box for the object and a second 2D bounding box generated from the second six DoF pose;
determine a total offset by combining the pose offset, the first projection offset, and the second projection offset; and
update parameters of a loss function based on the total offset and provide the updated parameters to the neural network to obtain an updated pose offset, updated first projections offset, and updated second projection offset that are then combined to obtain an updated total offset.