A 3D point cloud object recognition method, system, storage medium and product

By adopting a metric learning framework and a dual-network embedded model in 3D point cloud recognition, the problem of model overfitting in point cloud small sample recognition is solved, and more efficient point cloud object recognition is achieved.

CN119007188BActive Publication Date: 2025-03-28QUANZHOU INST OF EQUIP MFG +1
2 Cites 0 Cited by

Patent Information

Application Number
CN202411484121.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-10-23
Publication Date
2025-03-28
Estimated Expiration
2044-10-23

AI Technical Summary

Technical Problem

The existing 3D point cloud small sample recognition method is difficult to adapt to the disordered and irregular distribution of point cloud data, resulting in overfitting the model and unable to effectively recognize point cloud small sample object.

Method used

Using the metric learning framework, point cloud data is divided into meta-training data sets and meta-test data sets, and a dual-network embedding model is built on the embedded network, including dynamic graph convolutional networks and global feature extraction networks to extract local and global features. At the same time, a representation enhancement algorithm is designed on the classifier to feature enhancement of prototype vectors and query vectors.

Benefits of technology

Through the improved metric learning framework and dual-network embedded model, the features of point cloud data can be effectively extracted, avoid model overfitting, and improve the accuracy of point cloud small sample object recognition.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present invention relates to the field of 3D object recognition, and in particular to a 3D point cloud object recognition method, system, storage medium and product, S1: constructing a point cloud data set; S2: dividing the point cloud data set into a non-overlapping meta-training data set and a meta-testing data set, wherein the meta-training data set includes B samples; S3: constructing a dual-network embedding model including a parallel dynamic graph convolutional network and a global feature extraction network; S4: inputting the B samples into the dual-network embedding model, and training to obtain a B*1024 embedding vector; S5: using a representation enhancement algorithm module to perform a weighted operation on the B*1024 embedding vector to obtain an enhanced prototype vector and an enhanced query vector; S6: respectively calculating the distance between the enhanced prototype vector and the enhanced query vector, and classifying the query sample corresponding to the enhanced query vector into the category where the nearest enhanced prototype vector is located; making each type of prototype vector approach the center of its same type of query vector, and making each type of query vector approach its prototype vector at the same time.
Need to check novelty before this filing date? Find Prior Art