A gated gaussian geometry fusion pure visual timing BEV three-dimensional detection method

By employing a gated Gaussian geometric fusion method, the problems of depth blur and temporal jitter in pure visual 3D target detection are solved, achieving highly robust detection of distant and weakly textured targets, and enhancing the temporal consistency and spatial discriminative power of the detection results.

CN122392043APending Publication Date: 2026-07-14CHONGQING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHONGQING UNIV OF POSTS & TELECOMM
Filing Date
2026-05-13
Publication Date
2026-07-14

AI Technical Summary

Technical Problem

Existing purely visual 3D target detection methods are prone to geometric representation errors in distant targets, occluded targets, or weakly textured regions. Furthermore, temporal fusion methods are computationally intensive or prone to local geometric detail decay, making it difficult to effectively model the global long-distance dependence of images and specific semantic information in different dimensions.

Method used

A gated Gaussian geometric fusion method is adopted, which enhances the feature map through a dual-axis attention mechanism, performs temporal fusion using a gated loop mechanism, and constructs Gaussian prior features through Gaussian implicit encoding and cross-attention mechanism to generate features carrying spatial location and confidence information for 3D detection.

Benefits of technology

It improves the robustness of detection for distant and weakly textured targets, enhances the temporal consistency and coherence of detection results, and significantly improves spatial discriminative power and unified semantic and geometric representation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122392043A_ABST
    Figure CN122392043A_ABST
Patent Text Reader

Abstract

This application provides a purely visual temporal BEV 3D detection method based on gated Gaussian geometric fusion. The method extracts features from acquired multi-view perspective images; it then uses a dual-axis attention mechanism to perform horizontal and vertical modeling of the extracted features, obtaining an enhanced 2D feature map, which is projected onto the BEV space to generate the BEV features of the current frame; a gated loop mechanism is used to temporally fuse the BEV features of the current frame and the BEV features of the previous frame to obtain temporal BEV features; Gaussian implicit encoding is applied to the temporal BEV features to obtain Gaussian prior features; a cross-attention mechanism is used to fuse the Gaussian prior features with the temporal BEV features, and gated attention weights are generated based on the fused features; the gated attention weights are used to weighted fuse the temporal BEV features to obtain target features; and the target features are input into a 3D target detection head for analysis and processing to obtain the 3D detection results, thus improving the accuracy and stability of 3D target detection.
Need to check novelty before this filing date? Find Prior Art