Lidar point cloud image completion method and system based on guided dynamic refinement

By employing cross-modal cross-attention feature enhancement and structure-guided dynamic iterative refinement methods, the problem of insufficient cross-modal modeling of sparse depth maps and RGB images is solved, achieving high-precision dense depth map reconstruction and improving boundary clarity and structural consistency.

CN122415704APending Publication Date: 2026-07-17EAST CHINA UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-23
Publication Date
2026-07-17

AI Technical Summary

Technical Problem

In 3D scene reconstruction, existing technologies lack the ability to guide cross-modal modeling between sparse depth maps and RGB images, and the feature fusion in the decoding stage is insufficient. This results in excessive smoothing or error propagation in the depth completion results at boundaries and areas of structural abrupt changes, making it difficult to meet the requirements of high-precision scene reconstruction.

Method used

We employ a cross-modal cross-attention feature enhancement and a structure-guided dynamic iterative refinement method. By implementing fine cross-modal guidance in the encoding stage, enhancing multi-scale feature fusion in the decoding stage, and performing structure-aware refinement in the output stage, we generate a high-precision dense depth map.

Benefits of technology

It effectively improves the accuracy, boundary clarity, and structural consistency of depth completion results, and outputs dense depth maps of targets with sharp object boundaries and extremely high structural consistency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122415704A_ABST
    Figure CN122415704A_ABST
Patent Text Reader

Abstract

本发明公开了基于引导动态细化的激光雷达点云图像补全方法及系统,该方法包括:S1、预处理空间对齐的RGB图像和稀疏深度图;S2、提取多尺度图像特征与多尺度深度特征;S3、以多尺度深度特征作查询,对应尺度图像特征作键值,跨模态引导注意力计算得引导增强的深度特征;S4、将当前引导增强的深度特征与上尺度深度特征自适应引导融合,经通道注意力权重自适应加权调制得解码特征,生成初始稠密深度图;S5、提取RGB图像结构引导特征,预测该初始稠密深度图像素邻域传播权重和边缘抑制权重,更新输出目标稠密深度图。本发明通过跨模态交叉注意力特征增强与基于结构引导的动态迭代细化,实现稀疏激光雷达深度的高精度稠密补全。
Need to check novelty before this filing date? Find Prior Art