Point Cloud Clustering via Semantic Virtual Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional point cloud data clustering methods based on 3D space information often incorrectly cluster adjacent objects as a single entity, leading to inaccurate recognition of distinct objects in autonomous driving environments.
Innovation Solution
A method and system that utilize semantic segmentation through a pre-learned deep learning network to assign classes to point cloud data, store data in virtual layers based on these classes, and cluster points within each layer, allowing for accurate separation of adjacent objects by generating a grid map and grouping adjacent points into clusters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional 3D space information-based clustering method is used, then real-time performance is maintained, but object recognition accuracy deteriorates when objects are adjacent
Solution Approach 1:
The patent segments the clustering process by creating multiple virtual layers based on semantic segmentation results. Each virtual layer corresponds to a specific object class (e.g., vehicle, pedestrian, building), and clustering is performed independently within each layer. This segmentation approach prevents adjacent objects of different classes from being incorrectly merged into a single cluster, thereby improving object recognition accuracy while maintaining computational efficiency through parallel processing of discrete layers.
2Measurement precision
If semantic segmentation through deep learning network is added, then clustering accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies preliminary action by performing semantic segmentation before the clustering process. A pre-trained deep learning network assigns class labels to each point in the point cloud data in advance, creating structured virtual layers. This preliminary classification enables subsequent clustering operations to work with pre-organized data, reducing computational complexity during the actual clustering phase and allowing for real-time performance despite the added semantic segmentation step.
3Reliability
If virtual layers are created for each class, then adjacent objects are correctly separated, but data structure complexity increases
Solution Approach 1:
The patent introduces an additional dimensional organization to the point cloud data by creating virtual layers based on semantic class labels. Instead of organizing points solely by spatial coordinates, the system adds a class-based dimensional layer, where each virtual layer contains points belonging to a specific object class. This dimensional transformation enables reliable separation of adjacent objects through class-based filtering while maintaining a manageable data structure through systematic point-to-layer assignment rules.
Data Source
AI summary
A method for clustering point cloud data includes the following steps of identifying a class of each point data of the point cloud data, the class assigned according to a semantic segmentation processing of the point cloud data, storing a plurality of point data of the point cloud data in virtual layers based on the class assigned to each point data, the virtual layers each associated with at least one class; and clustering the plurality of point data for each of the virtual layers.