基于多尺度特征融合和边缘增强的多传感器融合3D目标检测算法

打开文本图片集

中图分类号：U469.79 文献标志码：A 文章编号：1000-582X（2025）08-078-08

doi：10.11835/j.issn.1000-582X.2025.08.007

Multi-sensor fusion 3D target detection algorithm based on multi-scale feature fusion and edge enhancement

LIU Jianguo 1，2 ，CHEN Wen¹²，ZHAO Yifan³，ZHOU Qi¹²，YAN Fuwul²，YIN Zhishuai1²， ZHENG Hao'，WU Youhua' （1.Foshan Xianhu Laboratory，Foshan，Guangdong 5282O0，P.R.China; 2.HubeiKey Laboratory ofAdvanced Technology for Automotive Components， Wuhan University of Technology， Wuhan 430070，P.R. China; 3. SAIC-GM-Wuling Automobile Co.， Ltd.， Liuzhou， Guangxi 545007， P.R. China）

Abstract： BEV （bird’s eye view）-based multi-sensor fusion perception algorithms for autonomous driving have made significant progressin recent years and continue to contribute to the development of autonomous driving. In the research of multi-sensor fusion perception algorithms，multi-view image-to-BEVconversion and multi-modal feature fusion have been the key chalenges in BEV perception algorithms.In this paper，we propose MSEPECRN，a fusion sensing algorithm of camera and millimeter-wave radar for 3D target detection， which utilizes edge featuresand pointclouds to improve theaccuracyof depth prediction，and thenrealizes theaccurate conversionof multi-view images to BEV features.Meanwhile，a multi-scale deformable large kernel attention mechanism is introduced for modal fusion to solve the misalignment problem due to theexcessive diference offeatures from different sensors.Experimental results on the nuScenes open-source dataset show that compared to the baseline network， the proposed algorithm achieves improvements of 2.17% in mAP， 1.93% in NDS， 2.58% in mATE， 8.08% in mAOE，and 2.13% in mAVE.This algorithm can effectively improve the vehicle’s ability to perceive movingobstacles on the road，and has practical value.

Keywords：3D target detection; bird’s eye view; multi-modal fusion; depth prediction

1多传感器融合的3D检测算法研究背景与方法概述

实现可靠的3D感知是自动驾驶的关键，是车辆在复杂多变环境中正常行驶的前提。（剩余10029字）