Domain adaptation point cloud pedestrian detection method

By employing voxelization processing, pseudo-label quality screening, cross-domain instance migration, and confidence-guided learning mechanisms, high-quality pseudo-labels are generated, solving the problem of insufficient pedestrian detection performance in underground scenes and achieving accurate pedestrian detection in underground scenes.

CN117935305BActive Publication Date: 2026-06-16TIANJIN UNIV
0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
TIANJIN UNIV
Filing Date
2024-01-10
Publication Date
2026-06-16

AI Technical Summary

⚠Technical Problem

Existing domain-adaptive point cloud detection methods struggle to generate pseudo-labels that balance quality and quantity in underground scenarios such as mines and tunnels due to significant inter-domain differences, thus impacting pedestrian detection performance.

⚗Method used

By extracting point cloud features through voxelization and sparse convolution, a pseudo-label quality screening strategy and a cross-domain instance transfer strategy are designed. A pseudo-label confidence-guided learning mechanism is constructed to optimize network learning, generate high-quality pseudo-labels, and perform self-training.

🎯Benefits of technology

It enables accurate detection of pedestrians in underground scenarios, mitigates the negative impact of low-quality false labels, and improves detection performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
Patent Text Reader

Abstract

The application discloses a domain adaptation point cloud pedestrian detection method, comprising the following steps: extracting features of voxelized point clouds through sparse convolution to obtain voxel feature representation of the point clouds, and sending the voxel features into a high compression module and a bird's eye view feature extraction module to obtain BEV features of the point clouds; dividing the quality of the pseudo labels of the target domain and removing low-quality pseudo labels; sampling instances from the source domain and migrating them to the underground scene of the target domain to obtain new pseudo labels of the target domain; constructing a pseudo label confidence guided learning mechanism to optimize network learning under the guidance of the pseudo label confidence; initializing the network by using the parameters obtained by pre-training the source domain, and optimizing the domain adaptation point cloud pedestrian detection network by using the pseudo label quality screening strategy, the cross-domain instance migration strategy and the pseudo label confidence guided learning mechanism. The application effectively obtains rich pseudo labels of the target domain, and guides network optimization by using the confidence of the pseudo labels, so that accurate pedestrian detection in the underground scene is realized.
Need to check novelty before this filing date? Find Prior Art