Urban 3D Point Cloud Building Extraction Without Point-by-Point Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing 3D point cloud processing methods, particularly for outdoor scenes, are time-consuming due to the large volume of data and the reliance on point-by-point clustering in bottom-up approaches, which affects the efficiency of building instance extraction.
Innovation Solution
A method involving the construction of a large-scale urban point cloud dataset that selects building candidate points from a foreground semantic map, determines a relational matrix based on instance perception features and center offset vectors, and scores the results to improve the accuracy and speed of building instance extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If point-by-point clustering is used to complete the instance of 3D point cloud, then the completeness of building instance extraction is improved, but the time efficiency and processing speed deteriorate due to large data volume
Solution Approach 1:
The patent segments the point cloud processing into two stages: first performs semantic segmentation to classify points into building and non-building categories, then performs instance segmentation only on the building points. This two-stage segmentation approach reduces the number of points requiring instance-level processing while maintaining extraction accuracy.
Solution Approach 2:
The patent applies semantic segmentation as a preliminary action before instance segmentation. By pre-classifying points into building and non-building categories, the system prepares the data in advance, allowing the subsequent instance segmentation to focus only on relevant building points rather than processing all points uniformly.
2Manufacturing precision
If point-by-point clustering is performed on outdoor scene 3D point cloud data, then the completeness of building instance extraction is improved, but the time consumption increases significantly
Solution Approach 1:
The patent divides the processing into semantic segmentation (classifying building vs. non-building points) followed by instance segmentation (grouping building points into instances). This segmentation allows the computationally intensive instance clustering to operate only on building points rather than all points in the scene.
Solution Approach 2:
Semantic segmentation is performed as a preliminary filtering step to identify and isolate building points from non-building points. This preliminary classification reduces the dataset size for the subsequent instance segmentation step, thereby reducing time consumption.
3Manufacturing precision
If all building foreground points are processed for instance extraction, then the accuracy is improved, but the processing complexity and time consumption increase
Solution Approach 1:
The patent segments the point cloud into building foreground points and non-building points through semantic segmentation, then performs instance extraction only on the building foreground points. This segmentation reduces processing complexity by eliminating irrelevant points from the instance extraction pipeline.
Solution Approach 2:
The patent extracts and isolates building foreground points from the complete point cloud dataset using semantic segmentation results. By taking out only the relevant building points for instance extraction, the system reduces processing complexity while maintaining accuracy for building instances.
Data Source
AI summary
A method and an apparatus for constructing large-scale urban point cloud dataset and extracting building instance. The method includes obtaining a point cloud feature corresponding to a target point cloud; determining a foreground semantic map, an instance perception feature, and a center offset vector corresponding to the target point cloud based on the point cloud feature; selecting a plurality of building candidate points from building foreground points in the foreground semantic map, determining a relational matrix between the building foreground points and the building candidate points based on the instance perception feature and the center offset vector, determining an instance tag of each building foreground point based on the relational matrix, and merging the instance tag to obtain a result of a building candidate instance; and scoring the result of the building candidate instance, and determining a result of a building instance based on the scoring.