Point Cloud Patch Matching for Neighbor-Aware Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing point cloud matching technologies based on supervised deep learning suffer from low matching accuracy due to the lack of consideration for neighboring points during feature extraction, requiring additional positioning sensors, increasing costs, and limiting offline data collection and out-of-the-box performance.
Innovation Solution
A point cloud matching method that divides a to-be-matched point cloud into patches, inputs these patches to a pre-trained feature coding module, and during pre-training, uses masked patches to improve feature extraction, ultimately acquiring a global description vector for matching against historical databases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If supervised deep learning algorithms (K nearest neighbor, three-dimensional sparse convolution) are used for point cloud matching, then the matching process can be automated, but the matching accuracy is low because neighboring point information is not considered in feature extraction
Solution Approach 1:
The point cloud is divided into multiple local patches, and feature extraction is performed separately for each patch. This segmentation allows the model to capture local geometric structures and neighboring point relationships that are lost in global feature extraction, thereby improving matching accuracy while maintaining automation.
Solution Approach 2:
Instead of treating the entire point cloud uniformly, the patent applies different processing to different regions by dividing it into patches. Each patch undergoes independent feature extraction that considers its local geometric context, enabling the system to preserve important local structural information for more accurate matching.
2Extent of automation
If additional positioning sensors are mounted to provide distance relationship markings for training data, then supervised deep learning can be implemented, but device costs increase
Solution Approach 1:
The system uses the point cloud data itself to provide the training supervision signal. By treating visible and masked patches in a self-supervised manner during training, the model learns to predict masked patch features from visible ones, eliminating the need for external positioning sensors while maintaining the benefits of supervised learning.
Solution Approach 2:
The point cloud processing pipeline serves multiple functions: it performs both the primary matching task and generates its own training data through the mask prediction mechanism. This multi-functionality allows the system to achieve supervised learning capabilities without requiring additional specialized sensors.
3Extent of automation
If supervised deep learning algorithms are used for point cloud matching, then feature extraction can be performed, but offline data collection and training is limited and out-of-the-box performance is not achieved
Solution Approach 1:
The model is pre-trained on large-scale datasets using the mask prediction objective, learning generalizable point cloud features before being applied to specific matching tasks. This preliminary training enables the system to achieve good out-of-the-box performance and adapt to different scenarios without requiring extensive task-specific data collection and fine-tuning.
4Productivity
If global features of point cloud are extracted by neural network, then point cloud matching can be achieved, but influence of neighboring points is not considered
Solution Approach 1:
The point cloud is divided into multiple local patches, and feature extraction is performed separately for each patch. This segmentation allows the model to capture local geometric structures and neighboring point relationships that are lost in global feature extraction, thereby improving matching accuracy while maintaining automation.
Solution Approach 2:
The patent transitions from global 3D space to local patch-level processing, adding a spatial hierarchy dimension. By processing patches independently and then aggregating their features, the system preserves local geometric information while maintaining the overall global context, effectively operating in multiple spatial dimensions simultaneously.
Data Source
AI summary
A point cloud matching method and apparatus, an electronic apparatus, and a storage medium are provided. The point cloud matching method includes: dividing a to-be-matched point cloud into a plurality of matched point cloud patches, and inputting the plurality of matched point cloud patches to a pre-trained point cloud feature coding module to obtain feature vectors of the plurality of matched point cloud patches; and acquiring a global description vector of the to-be-matched point cloud according to the feature vectors of the plurality of matched point cloud patches, matching the global description vector of the to-be-matched point cloud with global description vectors of point cloud frames in a preset historical database, and determining a point cloud frame in the historical database within a preset matching threshold range to be a point cloud matching result.