A pose correction-based fusion 3D target detection method and system

By generating high-quality pseudo-point clouds through local image depth estimation and view cone spatial pose correction algorithms, the problem of insufficient accuracy in 3D target detection in complex scenes is solved, achieving more stable and accurate detection results. It is suitable for low-cost autonomous driving systems that integrate LiDAR and vision cameras.

CN122157186APending Publication Date: 2026-06-05XIDIAN UNIV HANGZHOU RES INST +1
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
XIDIAN UNIV HANGZHOU RES INST
Filing Date
2026-01-19
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

Existing 3D object detection methods lack accuracy in complex scenes, especially when point clouds are sparse, targets are occluded, or detection is at a long distance. Furthermore, depth estimation errors in monocular images lead to noise and unclear structure in the reconstructed point cloud, affecting the accuracy and stability of 3D bounding box prediction.

Method used

An initial pseudo-point cloud is generated by local image depth estimation. The orientation error and nonlinear depth error are reduced by combining the pose correction algorithm in the view frustum space. A discrete pseudo-point cloud filtering algorithm with instance segmentation mask and adaptive sphere radius filtering strategy is adopted to remove noise points and improve the point cloud quality.

Benefits of technology

It achieves more stable and accurate 3D target detection in complex scenarios, significantly improves robustness and spatial consistency, reduces dependence on high-cost LiDAR, is suitable for scenarios where low-cost LiDAR and vision camera are fused, and improves detection performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122157186A_ABST
    Figure CN122157186A_ABST
Patent Text Reader

Abstract

The application belongs to the technical field of 3D target detection, and discloses a fusion 3D target detection method and system based on pose correction, which performs 2D target detection, local cutting and local mask segmentation on an RGB image, uses local image depth estimation to obtain an initial pseudo point cloud through local depth back projection, reduces orientation error and nonlinear depth error of the structured pseudo point cloud through a pose correction algorithm based on a view cone space, and removes long tail noise points caused by monocular depth estimation and inaccurate mask boundaries through a discrete pseudo point cloud filtering algorithm that fuses instance segmentation masks and an adaptive ball radius filtering strategy. Through efficient fusion of point cloud and image information, the application can realize high-precision and low-cost 3D target detection, effectively improves overall detection precision and small target recognition capability, and has good engineering practicability and expansion potential.
Need to check novelty before this filing date? Find Prior Art