LiDAR Video Labeling Using Frame-Difference Pixel Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing methods for labeling autonomous driving data require significant time and labor, and it is difficult to quickly process LiDAR point clouds and distinguish dynamic from static objects due to the manual labeling process.
Innovation Solution
A method and apparatus that generate and provide a data video for training an AI model by converting LiDAR point cloud data into 2D images, creating a video sequence, and allowing users to label objects through a user interface, enabling efficient and accurate labeling of dynamic and static objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling of LiDAR point cloud data is performed, then labeling accuracy can be achieved, but time consumption and labor costs increase significantly
Solution Approach 1:
The patent creates 2D projected images from 3D LiDAR point cloud data, generating visual copies that preserve spatial information while enabling faster processing. These 2D projections serve as intermediate representations that maintain the essential geometric relationships of the original 3D data, allowing annotators to work with simplified visual representations rather than raw point clouds.
Solution Approach 2:
The patent introduces 2D projected images as an intermediary format between the original 3D LiDAR data and the final labeling results. This intermediate representation facilitates easier visualization and annotation while preserving the necessary spatial information, acting as a bridge that reduces the complexity of direct 3D point cloud labeling.
2Measurement precision
If manual labeling of LiDAR point cloud data is performed, then object classification can be achieved, but labor costs increase
Solution Approach 1:
The patent generates 2D projected images as visual copies of the 3D point cloud data, creating simplified representations that maintain spatial relationships. These projected images enable annotators to perform object classification more efficiently while preserving the accuracy needed for distinguishing dynamic and static objects, thereby reducing labor costs.
Solution Approach 2:
The patent replaces the mechanical process of directly labeling 3D point cloud data with a system that uses 2D projected images. This substitution transforms the labeling task from working with complex three-dimensional coordinate data to working with two-dimensional visual representations, significantly reducing the cognitive and manual effort required.
3Loss of information
If LiDAR point cloud data is processed by drawing and graphing points individually, then complete data representation can be achieved, but processing speed decreases
Solution Approach 1:
The patent transforms 3D LiDAR point cloud data into 2D projected images by changing the dimensional representation. This dimensionality reduction projects three-dimensional spatial information onto a two-dimensional plane, preserving essential geometric relationships and spatial distribution while enabling much faster processing and visualization compared to handling individual 3D points.
Solution Approach 2:
The patent creates 2D projected copies of the 3D point cloud data that maintain the spatial distribution and relative positions of points. These projected images serve as efficient representations that preserve the essential structural information of the original data while enabling rapid processing and analysis.
4Ease of operation
If only one LiDAR point cloud is used for labeling, then labeling simplicity can be maintained, but object attribute determination becomes difficult
Solution Approach 1:
The patent transforms multiple LiDAR point clouds into a temporal video sequence by creating 2D projected images at different time points. This dimensional transformation from spatial 3D data to temporal video data enables annotators to observe object motion and attribute changes over time, significantly improving the ability to distinguish dynamic from static objects while maintaining operational simplicity through familiar video playback interfaces.
Solution Approach 2:
The patent performs preliminary processing by generating 2D projected images from multiple LiDAR point clouds before the labeling process. This pre-processing step creates a temporal sequence that preserves motion information and object attribute changes, enabling annotators to make more accurate determinations about object dynamics without having to manually analyze multiple 3D point clouds.
Data Source
AI summary
Provided are a method and apparatus for generating and providing a data video for generating training data of an artificial intelligence model and a recording medium on which a program for the same is recorded. According to various embodiments of the present disclosure, a method includes generating light detection and ranging (LiDAR) images using LiDAR point cloud data for a predetermined area, generating a LiDAR video using each of the LiDAR images as a unit frame, and providing a user interface (UI) that outputs the generated LiDAR video, when a unit frame change request is acquired from a user, comparing a first unit frame currently being output with a second unit frame after the first unit frame to detect at least one pixel whose attribute changes, and updating only the one or more detected pixels in the first unit frame using the second unit frame.