Title:
AERIAL VIEW-BASED MULTI-VIEW 3D OBJECT DETECTION METHOD, MEMORY AND SYSTEM
Document Type and Number:
WIPO Patent Application WO/2023/216460
Kind Code:
A1
Abstract:
An aerial view-based multi-view 3D object detection method, a memory and a system. The method comprises: by using a residual network and a feature pyramid, encoding a multi-view picture to obtain multi-scale features; by means of a mapping relation, mapping the multi-scale features to an aerial view to obtain aerial view features; performing random initialization on a query vector, constructing a plurality of sub-spaces by means of a first multi-head attention mechanism module, and projecting the query vector into the plurality of sub-spaces to obtain initialization features; performing first residual connection and normalization on the initialization features; by using a second multi-head attention mechanism module, combining the features subjected to the first residual connection and normalization with the aerial view features to obtain learning features; and performing second residual connection and normalization on the learning features, outputting an object detection category by using a first feedforward neural network module, and outputting the size of an object detection box by using a second feedforward neural network module.
More Like This:
Inventors:
CHEN YUANPENG (CN)
ZHANG JUNLIANG (CN)
ZHAO TIANKUN (CN)
ZHANG JUNLIANG (CN)
ZHAO TIANKUN (CN)
Application Number:
PCT/CN2022/114418
Publication Date:
November 16, 2023
Filing Date:
August 24, 2022
Export Citation:
Assignee:
HOZON NEW ENERGY AUTOMOBILE CO LTD (CN)
International Classes:
G06V20/56
Foreign References:
| CN114821505A | 2022-07-29 | |||
| CN113011317A | 2021-06-22 | |||
| CN113658100A | 2021-11-16 | |||
| CN111832655A | 2020-10-27 | |||
| CN114218999A | 2022-03-22 | |||
| CN113610044A | 2021-11-05 | |||
| US20210390714A1 | 2021-12-16 |
Attorney, Agent or Firm:
SHANGHAI PATENT & TRADEMARK LAW OFFICE, LLC (CN)
Download PDF: