Feature mining method and device, model training method and device, equipment, medium

By using image attention encoders and point cloud voxelization technology, the features of unmanned vehicle images and point clouds are automatically mined, which solves the problem of inaccurate positioning of unmanned vehicles on roads lacking high-precision map features, and improves the accuracy and efficiency of positioning.

CN117726995BActive Publication Date: 2026-06-23BEIJING BAIDU NETCOM SCI & TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
BEIJING BAIDU NETCOM SCI & TECH CO LTD
Filing Date
2023-12-06
Publication Date
2026-06-23

AI Technical Summary

Technical Problem

Existing technologies struggle to obtain accurate and reliable positioning results in autonomous vehicle driving, especially on unconventional roads lacking high-precision map features. Manual annotation is costly, and algorithm detection accuracy is low, leading to inaccurate positioning.

Method used

By inputting the image to be detected into a pre-trained image attention encoder, attention features are obtained, and the point cloud is converted into a two-dimensional detection grid. The visual mining features are determined using the bird's-eye view attention features, and a loss function is constructed for backpropagation to adjust the parameters of the image attention encoder, thereby achieving automated feature mining.

Benefits of technology

It reduces the human cost of manual feature mining, improves feature mining efficiency, obtains hidden visual features, and improves the accuracy and efficiency of unmanned vehicle positioning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117726995B_ABST
    Figure CN117726995B_ABST
Patent Text Reader

Abstract

The present disclosure provides feature mining method and device, model training method and device, equipment and medium, relates to the image processing technical field, especially to the automatic driving, artificial intelligence, deep learning and other technical fields. The specific implementation scheme is: input at least one to-be-detected image into a pre-trained image attention encoder to obtain an attention feature corresponding to the to-be-detected image; body voxelize a plurality of to-be-detected point clouds in a preset range corresponding to the to-be-detected image into a two-dimensional detection grid of a preset size, the two-dimensional detection grid comprises a plurality of detection grids, and each detection grid comprises at least one to-be-detected point cloud; determine a bird's eye view attention feature corresponding to the detection grid according to a projection relationship between the two-dimensional detection grid and the to-be-detected image and the attention feature; and determine a visual mining feature in the to-be-detected image according to the bird's eye view attention feature.
Need to check novelty before this filing date? Find Prior Art