3D Voxel Object Detection Refinement Network

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current object detection solutions, particularly in 3D scenarios, suffer from low accuracy due to the loss of positioning information during feature extraction and processing.

Innovation Solution

The method involves obtaining a point cloud of a scene, mapping it to a 3D voxel representation, performing convolution operations to extract feature information, and adjusting initial positioning information based on location and convolution features to improve detection accuracy by fusing location and convolution feature information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If point cloud data is mapped to 3D voxel representation for feature extraction, then object detection efficiency is improved, but positioning information is lost

Engineering Contradiction:
Improveobject detection efficiencyVSAvoidpositioning information
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent introduces a refinement network that operates in the original point cloud space (3D spatial coordinates) to compensate for information loss during voxelization. By adding this refinement dimension after the main voxel-based detection pipeline, the system recovers precise positioning information while maintaining the efficiency benefits of voxel representation.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The refinement network acts as an intermediary component between the voxel-based detection system and the final output. It takes the candidate object regions from the voxel pipeline and refines their positioning by incorporating detailed point cloud features, thus mediating between efficiency-oriented voxel processing and precision-oriented location accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Speed

If 3D voxel representation is used for object detection, then processing speed is improved, but detection accuracy deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoiddetection accuracy
Core Design Contradiction:
SpeedVSMeasurement precision

Solution Approach 1:

The patent segments the object detection process into two distinct stages: a coarse detection stage using efficient voxel representation for speed, and a refinement stage using detailed point cloud features for accuracy. This segmentation allows each stage to optimize for its specific goal without compromising the other.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The voxel-based detection performs preliminary action by quickly identifying candidate object regions and their approximate locations. This preliminary detection narrows down the search space, allowing the refinement network to focus computational resources only on relevant regions, thus achieving both speed and accuracy.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If feature extraction is performed on point cloud data, then object detection capability is improved, but positioning precision is reduced

Engineering Contradiction:
Improveobject detection capabilityVSAvoidpositioning precision
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The refinement network implements feedback by taking the candidate regions from feature extraction and using them to guide further processing. It feeds back detailed point cloud features specifically for the regions of interest, allowing the system to maintain high detection capability while recovering positioning precision through targeted refinement.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3955158B1Object detection method and apparatus, electronic device, and storage medium
Publication Date: 2024.05.01 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3955158B1 patent drawingFigure 1a
  • EP3955158B1 patent drawingFigure 1b
  • EP3955158B1 patent drawingFigure 1c~1d

AI summary

An object detection method and apparatus, an electronic device and a storage medium. Said method comprises the following steps: acquiring a point cloud of a scene, the point cloud comprising position information of points (101); mapping the point cloud to a three-dimensional voxel representation, so as to obtain feature information concerning a three-dimensional voxel (102); performing a convolution operation on the feature information concerning the three-dimensional voxel, to obtain a convolution feature set (103); determining, on the basis of the convolution feature set, initial positioning information concerning a candidate object area (104); selecting, from the point cloud, a target point located within the candidate object area (105); and on the basis of position information concerning the target point and target convolution feature information, adjusting the initial positioning information concerning the candidate object area, to obtain positioning information concerning a target object area, the target convolution feature information being a convolution feature in the convolution feature set corresponding to the position of the target point (106). The method above improves the accuracy of object detection.