Graph Neural Network 3D Object Tracking

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for tracking objects in a scene, particularly for autonomous vehicles and robots, are limited in providing reliable information about object positions and trajectories, especially in 3D environments, and struggle to distinguish between true and false detections.

Innovation Solution

A computer-implemented method using a graph neural network to calculate object information by iteratively acquiring detections, determining embedded features, updating track edges and nodes, assessing reliability, and generating new tracks, which includes the use of predictive trackers and neural networks to improve tracking accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If graph neural network methods are used for 2D object tracking, then information reliability is improved, but tracking accuracy in 3D environments deteriorates

Engineering Contradiction:
Improveinformation reliabilityVSAvoidtracking accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent transitions from 2D graph neural network tracking to 3D object tracking by introducing depth information and spatial transformations. The system processes 3D point cloud data and transforms detections into 3D space, enabling accurate trajectory estimation in three-dimensional environments while maintaining the reliability benefits of graph-based methods.

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

Solution Approach 2:

The patent changes the parameter space from 2D coordinates to 3D coordinates by incorporating depth information, vertical position, and spatial orientation. This parameter expansion allows the system to maintain high reliability through graph neural networks while achieving accurate 3D tracking, resolving the contradiction between reliability improvement and tracking accuracy deterioration.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If multiple detections are processed to improve tracking reliability, then false detections increase, but distinguishing true from false detections becomes more difficult

Engineering Contradiction:
Improvetracking reliabilityVSAvoiddetection verification difficulty
Core Design Contradiction:
ReliabilityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent implements feedback mechanisms through the graph neural network that continuously evaluates detection quality and updates tracking states. The system uses past track information to assess current detections, providing feedback that helps distinguish true detections from false ones, thereby maintaining high reliability while reducing the difficulty of detection verification.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent performs preliminary actions by pre-processing detections through the graph neural network before final track assignment. The system evaluates detection reliability in advance using graph-based relationships and historical track data, filtering out likely false detections before they contaminate the final tracking results, thus maintaining reliability while simplifying the verification process.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If iterative graph updating is performed to improve trajectory accuracy, then computational complexity increases, but real-time processing capability deteriorates

Engineering Contradiction:
Improvetrajectory accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies partial updating strategies where not all graph nodes and edges are updated in every iteration. The system selectively updates only the necessary portions of the graph based on current detection quality and track stability, reducing computational complexity while maintaining high trajectory accuracy through targeted iterative refinement.

Inventive Principle:
Principle #16Partial or excessive action

4Reliability

If 3D spatial information is incorporated to improve trajectory reliability, then data processing requirements increase, but measurement precision deteriorates

Engineering Contradiction:
Improvetrajectory reliabilityVSAvoidposition precision
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent transforms 3D spatial data into meaningful parameters for trajectory estimation by incorporating depth, vertical position, and spatial orientation into the graph neural network. This parameter transformation enables the system to process 3D information efficiently while maintaining high position precision, resolving the contradiction between reliability improvement and precision deterioration.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4050510A1Object information calculation method and system
Publication Date: 2022.08.31 TOYOTA JIDOSHA KK
  • EP4050510A1 patent drawingFigure 1
  • EP4050510A1 patent drawingFigure 2
  • EP4050510A1 patent drawingFigure 3

AI summary

An object information calculation method or system for obtaining object information relative to object(s). A graph (G) is a data structure comprising detection nodes (VD,i), track nodes (VT,k), detection edges (EDD,ij) and track edges (ETD,km). The method comprises performing iteratively: S30) acquiring detection(s) (DET) of object(s) so as to prepare base vectors (x); S40) determining embedded features (eD,i, eT,m, eDD,ij, eTD,mk) of the object(s) for nodes and edges of the graph (G); S50) determining initial values (h(0)D,i, h(0)T,m, h(0)DD,ij, h(0)TD,mk) of the nodes and edges of the graph (G); S60) updating one or more time (i=1...L) at least track edge(s) (h(s)TD,mk) of the graph (G); S70) assessing a reliability of edge(s) and/or node(s), S80) updating track(s); S90) generating new track(s); S100) updating track node(s); S110) assessing whether a stop criterion has been fulfilled, and when the stop criterion has been fulfilled: S120) outputting object information relative to the object(s).