A vehicle target detection method based on radar vision fusion

By combining adaptive radar point cloud enhancement, multi-pooling point cloud feature encoding, and a shareable multi-semantic space attention module, the problem of insufficient detection accuracy of radar-camera fusion methods for long-distance and small-scale targets under sparse point cloud conditions is solved, achieving higher detection accuracy and robustness.

CN122176691APending Publication Date: 2026-06-09XIAN UNIV OF SCI & TECH
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
XIAN UNIV OF SCI & TECH
Filing Date
2026-05-11
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

Existing radar-camera fusion methods struggle to effectively detect distant and small-scale targets when dealing with sparse radar point clouds in complex traffic scenarios, resulting in insufficient detection accuracy.

Method used

An adaptive radar point cloud enhancement module (ARHGM) is used to generate hybrid point clouds. Virtual points are generated through Gaussian and uniform distributions, and the number ratio of virtual points is dynamically adjusted. The feature extraction capability is improved by combining a multi-pooling point cloud feature encoding module (MultiPool) and a shareable multi-semantic space attention module (SMSA), thereby enhancing the point cloud density and reliability.

Benefits of technology

It significantly improves the detection accuracy of distant and small-scale targets under sparse point cloud conditions, reduces missed detections, and enhances the quality and detection accuracy of mixed point clouds, especially showing better detection performance in complex traffic scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122176691A_ABST
    Figure CN122176691A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of target detection technology and discloses a vehicle target detection method based on radar vision fusion. The method includes generating an instance mask from an image, projecting the original point cloud onto an image plane, and using points falling within the instance mask area as foreground points. A circular area centered on the foreground points and with a preset radius is defined as a first virtual point generation area, and an area located outside the instance mask area and the first virtual point generation area is defined as a second virtual point generation area. First virtual points are generated in the first virtual point generation area using a Gaussian distribution; second virtual points are generated in the second virtual point generation area using a uniform distribution; the first and second virtual points are mixed and projected back to the radar coordinate system, and merged with the foreground points to obtain a mixed point cloud. This invention effectively solves the problem of insufficient detection performance caused by sparse radar point clouds and loss of feature encoding information, significantly improving the detection accuracy of long-range and small-scale targets.
Need to check novelty before this filing date? Find Prior Art