公开(公告)号：US20240029422A1

公开(公告)日：2024-01-25

申请号：US17871335

申请日：2022-07-22

Applicant: Robert Bosch GmbH

Inventor： Ehsan QASEMI ,  Alessandro OLTRAMARI

IPC: G06V10/94 ,  G06V10/764 ,  G06V20/40 ,  G06V20/52 ,  H04N7/18 ,  G06V10/86 ,  G06F40/20

CPC classification number: G06V10/945 ,  G06V10/764 ,  G06V20/41 ,  G06V20/52 ,  H04N7/183 ,  G06V10/86 ,  G06F40/20 ,  G06V2201/07 ,  G08G1/0125

Abstract: A human-assisted neuro-symbolic system for outputting fine-grained classifications and corresponding images or video of a desired object or scene. The system includes one or more cameras configured to generate a video feed of a scene. One or more processors are programmed to generate video analytics data from the video feed, including coarse-grained classification data regarding one or more objects in the scene. A knowledge graph is built with instantiated (e.g., time-based) domain ontology of the one or more objects in the scene. The domain ontology can be augmented via human-in-the-loop. Once augmented, the knowledge graph can be infused into a deep learning model, such as a natural language model. An input (e.g., in natural language) can seek fine-grained input characteristics, and the deep learning model infused with the knowledge graph retrieves a corresponding portion of the video feed with the fine-grained input characteristics.