3D Point Cloud Mapping for Single-Sensor Object Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for 3D object classification using Lidar data are limited by the need for multiple sensors, hand-crafted feature selection, and poor performance in varying environmental conditions, such as night or rainy weather, due to the inability to adapt to unique situations and low recognition accuracy for objects at far distances.

Innovation Solution

A processing method that transforms 3D point cloud data from a single depth sensor into a single multi-channel 2D space combining intensity, height, and distance information, which is well-suited for machine learning algorithms, allowing deep neural networks to automatically determine feature combinations and improve classification accuracy across different conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If hand-crafted feature selection is used for 3D object classification, then the processing pipeline is simplified and can be tuned for specific situations, but the system cannot adapt to unique or rare situations and pre-filters significant information

Engineering Contradiction:
Improveease of implementing feature selectionVSAvoidadaptability to unique situations
Core Design Contradiction:
Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The deep neural network automatically performs feature selection and extraction from raw 3D point cloud data without requiring hand-crafted features. The system serves itself by learning optimal features directly from data, enabling adaptation to unique situations while eliminating the need for manual feature engineering

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system transforms the fixed hand-crafted feature approach into a dynamic parameter-based solution where the neural network learns optimal feature representations. By changing from static hand-crafted parameters to dynamic learned parameters, the system gains adaptability to various situations including rare cases

Inventive Principle:
Principle #35Parameter changes

2Loss of information

If multiple sensors (camera and depth sensor) are combined for 3D object classification, then more information is available for classification, but the system complexity increases and integration of different data types is required

Engineering Contradiction:
Improveinformation availability for classificationVSAvoidsensor integration complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent extracts and utilizes only the essential intensity and 3D coordinate information from the depth sensor, discarding the need for separate camera RGB data. This extraction of key features from a single sensor reduces system complexity while maintaining classification effectiveness

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The depth sensor is made multi-functional by extracting both 3D spatial coordinates and intensity information from the same sensor, eliminating the need for separate camera and depth sensor. This universal approach reduces device complexity while preserving necessary information

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Ease of operation

If intensity data is processed as 2D image maps for object recognition, then processing is simplified, but parallel access is limited and only single source of information is utilized

Engineering Contradiction:
Improveprocessing simplicityVSAvoidinformation access limitation
Core Design Contradiction:
Ease of operationVSLoss of information

Solution Approach 1:

The patent transforms the processing approach by organizing data into a multi-channel 2D representation where each channel preserves a different aspect (intensity, x-coordinate, y-coordinate, z-coordinate). This dimensional transformation maintains processing simplicity while enabling parallel access to multiple information sources simultaneously

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

4Loss of time

If feature pre-selection is performed before deep learning, then processing time is reduced, but deep neural networks cannot realize their potential in automatically finding features

Engineering Contradiction:
Improveprocessing time efficiencyVSAvoidautomatic feature discovery capability
Core Design Contradiction:
Loss of timeVSExtent of automation

Solution Approach 1:

The patent performs preliminary organization of raw 3D point cloud data into a structured multi-channel 2D format with intensity, x, y, and z channels. This preliminary action prepares data for efficient processing while preserving all necessary information, enabling both fast processing and automatic feature discovery by the neural network

Inventive Principle:
Principle #10Preliminary action

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

This approach enhances classification performance by enabling deep neural networks to learn stable patterns from combined intensity, height, and distance information, improving recognition accuracy and reaction time, especially in challenging environmental conditions, and reduces the need for multiple sensors, resulting in a more economical and efficient obstacle detection system for autonomous vehicles.

Implementation Method 1

A Lidar sensor measures distance by illuminating a target with a laser light

Methodology Applied
Scientific EffectLIDAR: LIDAR

Implementation Method 2

illuminating a target with a laser light

Methodology Applied
Scientific EffectLaser: Laser

Data Source

PatentEP3430427B1Processing method of a 3D point cloud
Publication Date: 2021.07.21 IMRA EURO
  • EP3430427B1 patent drawingFigure 1
  • EP3430427B1 patent drawingFigure 2
  • EP3430427B1 patent drawingFigure 3A~3B

AI summary

A processing method of a three-dimensional (3D) point cloud, comprising the steps of: obtaining a 3D point cloud from a predetermined view point of a depth sensor; extracting 3D coordinates (xyz) and intensity (i) data from each point of the 3D point cloud with respect to the view point; transforming 3D coordinates (xyz) and intensity (i) data into at least three two-dimensional (2D) spaces, namely an intensity 2D space (IS) function of the intensity (i) data of each point, a height 2D space (HS) function of an elevation (h) data of each point, and a distance 2D space (DS) function of a distance (d) data between each point of 3D point cloud and the view point, defining a single multi-channel 2D space (IS, HS, DS).