Lidar Point Cloud Segmentation for Automated Vehicles

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for point-cloud segmentation in lidar data for autonomous vehicles are computationally demanding and inefficient, particularly due to the irregular access of points in three-dimensional space, which slows down the processing of unorganized point-clouds and does not effectively exploit layer-based organization.

Innovation Solution

A methodology that organizes lidar point-clouds in layers with elliptical structures, using smart indexing and a two-step process involving ground plane extraction and clustering of non-ground points through the Ground Plane Fitting and Scan Line Run algorithms, which efficiently segments the data by treating points as pixels in an image-like structure.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional clustering algorithms traverse the point-cloud in an irregular way to find neighbors, then complete segmentation can be achieved, but processing speed significantly decreases due to exhaustive search

Engineering Contradiction:
Improvesegmentation completenessVSAvoidprocessing speed
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments the point-cloud data by first separating ground points from non-ground points using plane fitting, then applies clustering only to the non-ground points. This segmentation reduces the search space for neighbor finding and eliminates exhaustive traversal of all points, thereby improving processing speed while maintaining segmentation completeness.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary ground plane extraction and point classification before applying clustering algorithms. By pre-organizing points into ground and non-ground categories and establishing a coordinate system aligned with the ground plane, the system avoids unnecessary computations during the clustering phase, thus improving processing speed without compromising segmentation accuracy.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If unorganized point-clouds are processed using irregular access methods, then all points can be analyzed, but computational complexity increases and processing efficiency decreases

Engineering Contradiction:
Improvepoint analysis completenessVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent transforms the problem from three-dimensional spatial search to a two-dimensional projection problem by projecting points onto the ground plane and working in a coordinate system aligned with the ground. This dimensionality reduction simplifies neighbor search and reduces computational complexity while maintaining complete point analysis capability.

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

Solution Approach 2:

The patent changes the coordinate system parameters by defining a new reference frame where the ground plane normal becomes the Z-axis. This parameter transformation simplifies the mathematical operations required for plane fitting and distance calculations, reducing computational complexity while enabling complete analysis of all points.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3293670B1Lidar object detection system for automated vehicles
Publication Date: 2019.05.15 APTIV TECHNOLOGIES LTD
  • EP3293670B1 patent drawingFigure 1A~1B
  • EP3293670B1 patent drawingFigure 1C~1D
  • EP3293670B1 patent drawingFigure 2A~3B

AI summary

An object-detection system (10) suitable for an automated vehicle includes a lidar (20) and a controller (26). The lidar (20) is used to detect a point-cloud (22) that is organized into a plurality of scan-lines. The controller (26) is in communication with the lidar (20). The controller (26) is configured to classify each detected point in the point-cloud (22) as a ground-point (28) or a non-ground-point (30), define runs (32) of non-ground-points (30), where each run (32) characterized by one or multiple instances of adjacent non-ground-points (30) in a scan-line (24) separated from a subsequent run (32) of one or more non-ground-points (30) by at least one instance of a ground-point (28), define a cluster of non-ground-points (30) associated with the object (14). The cluster is characterized by a first run from a first scan-line being associated with a second run from a second scan-line when a first point from the first run is displaced less than a distance-threshold (36) from a second point from the second run.