A real-time point cloud semantic segmentation method based on attention mechanism and sparse tensor

By optimizing the neural network structure with sparse grids and attention modules, the problems of high computational cost and insufficient accuracy of existing methods are solved, enabling real-time point cloud semantic segmentation on low-computing-power platforms and improving feature extraction and prediction accuracy.

CN116229060BActive Publication Date: 2026-06-26DALIAN MARITIME UNIVERSITY
View PDF 3 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
DALIAN MARITIME UNIVERSITY
Filing Date
2023-01-04
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

Existing semantic segmentation methods cannot achieve a balance between high accuracy and real-time performance in mobile robot systems. Traditional convolutional neural networks consume a lot of computation and waste resources when processing 3D point cloud data, and existing methods are difficult to meet the requirements of real-time prediction.

Method used

A sparse grid computing method is adopted to reduce the computational load, and the neural network structure is optimized by combining spatial attention module and channel attention module. Real-time point cloud semantic segmentation is achieved through sparse tensors, thereby improving the accuracy of feature extraction and prediction.

Benefits of technology

Real-time computation and good prediction accuracy were achieved on a low-computing-power platform, solving the problems of large computational load and insufficient accuracy, and achieving a good balance between accuracy and prediction speed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116229060B_ABST
    Figure CN116229060B_ABST
Patent Text Reader

Abstract

The application discloses a real-time point cloud semantic segmentation method based on an attention mechanism and a sparse tensor, and comprises the following steps: acquiring point cloud data, forming a point cloud data set, and dividing the point cloud data set into test set data; preprocessing the point cloud data to obtain preprocessed point cloud data; performing data enhancement on the preprocessed point cloud data to obtain data-enhanced point cloud data, and dividing the data-enhanced point cloud data set into a training set of the point cloud data after reprocessing; establishing a semantic segmentation network model for semantic segmentation, training the semantic segmentation network model based on the training set of the data-enhanced point cloud data, evaluating the semantic segmentation network model by using a cross-validation set, and obtaining a trained semantic segmentation network model; and based on the test set data, the trained semantic segmentation network model is used to realize semantic segmentation of the point cloud data. The method realizes real-time operation on a low-computing-power platform and has good prediction accuracy.
Need to check novelty before this filing date? Find Prior Art