US 12,469,245 B2
Systems and methods for performing computer vision task using a sequence of frames
Manpreet Singh Takkar, Brampton (CA); Zixuan Hu, Markham (CA); and Wei Zhou, Richmond Hill (CA)
Assigned to HUAWEI TECHNOLOGIES CO., LTD., Shenzhen (CN)
Filed by Manpreet Singh Takkar, Brampton (CA); Zixuan Hu, Markham (CA); and Wei Zhou, Richmond Hill (CA)
Filed on Oct. 21, 2021, as Appl. No. 17/507,489.
Claims priority of provisional application 63/225,787, filed on Jul. 26, 2021.
Prior Publication US 2023/0033548 A1, Feb. 2, 2023
Int. Cl. G06V 10/22 (2022.01); G06V 10/10 (2022.01)
CPC G06V 10/22 (2022.01) [G06V 10/16 (2022.01)] 20 Claims
OG exemplary drawing
 
9. A computing system comprising:
a processing unit configured to execute instructions to cause the computing system to:
obtain a current frame corresponding to a current timestep in a sequence of frames, and a previous frame corresponding to a previous timestep in the sequence of frames that is previous to the current timestep, wherein the previous frame contains a tracked object that has been previously tracked in the previous frame and wherein the tracked object is present in the current frame and a current bounding box for the tracked object is absent in the current frame;
obtain a stored bounding box for the tracked object in the previous frame;
define a region of interest (ROI) in the previous frame, the ROI being defined based on the stored bounding box;
crop the current frame and the previous frame to the defined ROI;
compute a differential image between the cropped current frame and the cropped previous frame; and
using the differential image to predict the current bounding box for the tracked object in the current frame by forward propagating the differential image through a neural network that is trained to perform an object tracking task by using the differential image as input to generate a predicted output including the current bounding box for the tracked object in the current frame.