LIDAR Obstacle Localization Using Multi-Clustering Point Clouds

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for positioning pedestrians and bicycles using laser radar in self-driving vehicles are inaccurate and fail to meet real-time requirements due to the low number of reflection points, often clustering these objects with interfering objects, leading to missed detections.

Innovation Solution

A multi-clustering method is employed to extract a target point cloud representing the surface features of pedestrians and bicycles, involving multiple iterations of clustering to ensure accuracy and real-time positioning by identifying the highest probability point clouds.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a single clustering method is used to position obstacles in point cloud data, then the positioning process is simple and fast, but the accuracy is insufficient for objects with few reflection points such as pedestrians and bicycles

Engineering Contradiction:
Improvepositioning accuracyVSAvoidclustering method complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the point cloud data processing into multiple clustering stages: first clustering to obtain initial obstacle candidates, then second clustering to refine and separate overlapping objects. This segmentation approach enables accurate positioning of pedestrians and bicycles by breaking down the complex single-step clustering into manageable sequential steps, thereby resolving the contradiction between accuracy and complexity.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If deep learning methods are used for point cloud processing to achieve high recognition rates, then the recognition accuracy improves, but the computing time increases and real-time requirements are not met

Engineering Contradiction:
Improverecognition accuracyVSAvoidcomputing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent replaces deep learning computational methods with traditional geometric clustering algorithms. By substituting the complex neural network computations with efficient geometric operations like distance calculations and point cloud clustering, the system achieves comparable recognition accuracy while meeting real-time processing requirements for vehicle-mounted systems.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Productivity

If non-neural network clustering methods are used for obstacle positioning, then the computing time is reduced and real-time processing is achieved, but the accuracy for objects with few reflection points deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidpositioning accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent segments the clustering process into multiple sequential stages with different objectives. The first clustering stage rapidly identifies potential obstacle regions, while the second clustering stage refines the positioning accuracy. This multi-stage segmentation maintains high processing speed while improving accuracy for objects with few reflection points, resolving the contradiction between productivity and precision.

Inventive Principle:
Principle #1Segmentation

4Reliability

If clustering is performed on the entire point cloud to position all obstacles, then comprehensive detection is achieved, but objects with fewer reflection points are mistakenly detected or missed

Engineering Contradiction:
Improvedetection reliabilityVSAvoiddetection accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent segments the detection process into multiple clustering passes. The first clustering provides comprehensive obstacle detection, while the second clustering refines the results by separating overlapping objects and reducing false positives. This segmented approach maintains high detection reliability while improving detection accuracy for objects with few reflection points.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements feedback by using the results of the first clustering as input for the second clustering. The second clustering refines the initial results by re-evaluating points in the context of already-identified obstacles, thereby reducing false detections and improving overall detection reliability and accuracy.

Inventive Principle:
Principle #23Feedback

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

Accurately and efficiently positions pedestrians and bicycles by reducing false positives and negatives, enhancing the safety and reliability of self-driving systems.

Implementation Method 1

the first point cloud is obtained according to a laser measurement principle

Methodology Applied
Scientific EffectLIDAR: LIDAR

Data Source

PatentEP3933439B1Localization method and localization device
Publication Date: 2025.12.10 YINWANG INTELLIGENT TECHNOLOGIES CO LTD
  • EP3933439B1 patent drawingFigure 1
  • EP3933439B1 patent drawingFigure 2
  • EP3933439B1 patent drawingFigure 3~4

AI summary

Embodiments of this application disclose a positioning method and apparatus. The positioning method and apparatus relate to the artificial intelligence field, and specifically to the self-driving field. The positioning method includes: clustering points in a first point cloud through multi-clustering to obtain a target point cloud, where the target point cloud is a point cloud that represents a feature of a target object, and the first point cloud includes the target point cloud and a point cloud that represents a feature of an interfering object; and determining a position of the target object based on the target point cloud. In the embodiments of this application, the point cloud that represents the feature of the target object is selected through multi-clustering, to achieve accurate and highly real-time positioning of an obstacle with a relatively small quantity of reflection points.