一种可变形注意力的三维点云目标检测方法、系统和设备
By employing a deformable attention method in 3D point cloud target detection, the grid point positions and feature extraction are adaptively adjusted, solving the problem of insufficient utilization of sparse features in existing methods and achieving higher-precision target detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2023-12-27
- Publication Date
- 2026-07-17
AI Technical Summary
Existing two-stage 3D target detection methods fail to fully utilize the prediction results of the previous stage and the 3D sparse features extracted by the sparse convolutional layer, making it difficult to obtain rich contextual information around the target, resulting in insufficient detection accuracy.
A deformable attention-based 3D point cloud target detection method is adopted. By extracting 3D voxel features, using post-processing algorithms to filter candidate boxes, generating regions of interest, and adaptively adjusting the position and features of grid points within the regions of interest, combined with multi-scale feature extraction, the position of candidate boxes is generated.
It improves the accuracy of 3D target detection, especially for small targets at long distances, by acquiring richer contextual features and thus enhancing detection accuracy.
Smart Images
Figure CN117710659B_ABST