US 11,948,327 B2
Method and system for joint object location and ground plane estimation in computer vision
Roman Akulshin, San Francisco, CA (US); and Sihao Ding, Sunnyvale, CA (US)
Assigned to Volvo Car Corporation, Gothenburg (SE)
Filed by Volvo Car Corporation, Gothenburg (SE)
Filed on Jun. 7, 2021, as Appl. No. 17/340,228.
Claims priority of provisional application 63/035,878, filed on Jun. 8, 2020.
Prior Publication US 2021/0383567 A1, Dec. 9, 2021
Int. Cl. G06T 7/73 (2017.01); G06T 7/11 (2017.01); G06T 7/80 (2017.01)
CPC G06T 7/73 (2017.01) [G06T 7/11 (2017.01); G06T 7/80 (2017.01); G06T 2207/10016 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/30261 (2013.01); G06T 2210/12 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method, comprising:
obtaining an image using one of a camera and a perception sensor;
detecting and disposing one or more bounding boxes around the object in the image;
generating a reference cube, wherein the reference cube is assumed to be disposed at a center of a coordinate system associated with the one or more of the camera and the perception sensor, and wherein the reference cube is a model to which a projection matrix associated with the one or more bounding boxes indicating rotation and translation in three dimensions is applied;
projecting corners of the reference cube to respective corners of the one or more bounding boxes;
calculating reference cube-to-object homographies for front and back faces of the one or more bounding boxes using a direct linear transformation;
performing nonlinear least squares optimization for the reference cube-to-object homographies;
recovering rotation angles and translation distances for the object and combining them to form final homographies for the front and back faces of the one or more bounding boxes around the object; and
applying an inverse of a camera or perception sensor calibration matrix to an average of the final homographies to recover a true rotation and translation of the object.