A road marking extraction and classification method based on a local shape transformer network

By explicitly modeling the local geometry and anisotropic shape cues of point clouds using a local shape Transformer network, the problems of fragmented road marking structures and class imbalance in LiDAR point clouds are solved, improving the extraction accuracy and boundary integrity of fine markings, especially the recognition accuracy of long-tailed categories.

CN122368949APending Publication Date: 2026-07-10HINTON SPACE-TIME INTELLIGENT INNOVATION RESEARCH INSTITUTE MINHANG DISTRICT SHANGHAI +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610374865.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-25
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing point cloud deep learning networks struggle to explicitly perceive local anisotropic shapes when processing LiDAR point clouds, resulting in overly smoothed road marking features, blurred boundaries, and poor generalization ability under class imbalance, especially for long-tailed categories.

Method used

A local shape-based Transformer network is adopted. Through a local feature aggregation module, a local shape perception module, and a feature adaptive fusion module, combined with a multi-task joint loss function, the local geometry and anisotropic shape cues of the point cloud are explicitly modeled, and multi-stage feature extraction and restoration are performed to output the road marking category prediction results.

Benefits of technology

It improves the extraction accuracy and boundary integrity of road marking structures in sparse point clouds, especially the extraction accuracy and boundary integrity of fine markings, significantly improves the recognition accuracy of long-tail categories, and enhances the robustness and generalization ability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122368949A_ABST
    Figure CN122368949A_ABST
Patent Text Reader

Abstract

This invention provides a method for road marking extraction and classification based on a local shape Transformer network. The method includes: acquiring an original LiDAR point cloud and performing non-ground point filtering preprocessing to obtain a road surface point cloud; inputting the roadside point cloud into a Transformer-based encoder network, and obtaining semantic features containing local geometric details and global shape context through multi-stage downsampling and feature extraction; using a decoder network to upsample and recover the semantic features, fusing multi-scale information with skip connections, and outputting the road marking category prediction result for each point through a classification head. This invention improves the extraction accuracy and boundary integrity of fine markings such as lane lines and arrows.
Need to check novelty before this filing date? Find Prior Art