Vehicle Posture Recognition Using Iterative Component Confidence

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing vehicle posture recognition methods in autonomous driving, such as PointNet, are slow in achieving accurate clustering due to processing global features, leading to long waiting times and difficulty in meeting real-time requirements.

Innovation Solution

A method that utilizes a feature map to determine vehicle components and their position relationships, iteratively refining component and vehicle plane recognition based on confidence levels, leveraging local features to quickly and accurately recognize vehicle posture.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If PointNet processes point cloud data using multilayer perceptron to achieve accurate clustering, then clustering accuracy is improved, but processing speed deteriorates leading to long waiting time

Engineering Contradiction:
Improveclustering accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent segments the point cloud processing into multiple iterations. In each iteration, only unclassified points are processed rather than the entire point cloud, dividing the global processing task into smaller local processing steps that accumulate to achieve complete classification with better efficiency

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary classification in each iteration by processing only the unclassified points with the current model predictions, rather than reprocessing all points. This preliminary action on subsets of data reduces overall processing time while maintaining accuracy through iterative refinement

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If global features are used for accurate clustering of point cloud data, then clustering precision is improved, but real-time performance deteriorates

Engineering Contradiction:
Improveclustering precisionVSAvoidwaiting time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent divides the global feature processing into iterative local processing steps. Each iteration processes only unclassified points with their local features, segmenting the computational workload to reduce waiting time while accumulating clustering precision through multiple passes

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial action by processing only a portion of the points (unclassified points) in each iteration rather than all points. This partial processing reduces computational burden and waiting time, while the iterative nature ensures eventual complete classification with high precision

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3968221B1Vehicle attitude recognition method and related device
Publication Date: 2026.03.11 YINWANG INTELLIGENT TECHNOLOGIES CO LTD
  • EP3968221B1 patent drawingFigure 1
  • EP3968221B1 patent drawingFigure 2~3
  • EP3968221B1 patent drawingFigure 4~5

AI summary

Disclosed are a vehicle posture recognition method and apparatus. A feature map represents probabilities that each feature point is located on different components of a target vehicle. A plurality of components of the target vehicle included in the feature map are determined based on a plurality of feature points in the feature map. A vehicle plane of the target vehicle is determined based on a position relationship between the plurality of components, and a posture of the target vehicle is further recognized based on the determined vehicle plane. The process of obtaining the vehicle plane includes a plurality of iterations. In an ith iteration, a component recognition result of a target feature point is determined based on probabilities that the target feature point is located on different components in an (i-1)th iteration. The component recognition result has a first result confidence. A vehicle plane recognition result is determined based on a position relationship between a plurality of to-be-determined components. The vehicle plane recognition result has a second result confidence. Probabilities that the target feature point is located on different components in the ith iteration are obtained by updating based on the first result confidence and the second result confidence.