基于置信估计网络和改进D-S的多模态融合方法  

Multimodal Fusion Method Based on Confidence Estimation Network and Improved D-S Theory

作　　者：程腾 郭利港 张强[1,2,3,4] 王文冲 石琴[1,2,3] 侯登超[1,2,3] CHENG Teng;GUO Ligang;ZHANG Qiang;WANG Wenchong;SHI Qin;HOU Dengchao(Key Laboratory for Automated Vehicle Safety Technology of Anhui Province,Hefei University of Technology,Hefei 230009,China;Engineering Research Center for Intelligent Transportation and Cooperative Vehicle-Infrastructure of Anhui Province,Hefei 230009,China;School of Automotive and Transportation Engineering,Hefei University of Technology,Hefei 230009,China;Chery Automobile Co.,Ltd.,Wuhu 241007,Anhui,China)

机构地区：[1]合肥工业大学自动驾驶汽车安全技术安徽省重点实验室,合肥230009 [2]安徽省智慧交通车路协同工程研究中心,合肥230009 [3]合肥工业大学汽车与交通工程学院,合肥230009 [4]奇瑞汽车股份有限公司,安徽芜湖241007

出　　处：《汽车工程学报》2025年第2期137-146,共10页Chinese Journal of Automotive Engineering

基　　金：国家自然科学基金项目(82171012);安徽省自然科学基金资助项目(2208085MF171);安徽省新能源汽车暨智能网联汽车创新工程项目(JZ2021AFKJ0002);汽车标准化公益性开放课题资助(CATARC-Z-2022-01350)。

摘　　要：针对神经网络可解释性较低,且多模态融合中D-S理论易在高冲突场景出现悖论的问题,提出了一种基于置信估计网络和改进D-S理论的结果级多模态融合方法,该方法包括一种置信估计网络,将目标检测中的分类问题表述为置信估计问题,输出单模态网络检测结果的置信估计分数;一种改进D-S理论的融合方法,用检测结果的置信度分数和类别信息合成证据,最终融合不同模态检测数据。在KITTI数据集上进行的评估试验表明,该融合方法的mAP值较图像单模态检测最高提升6.64%,较点云单模态检测最高提升15.43%,对比结果表明,该融合方法的mAP提升值高于经典D-S融合0.81%,能有效减少分类冲突,克服经典D-S的局限性。Neural networks lack interpretability and the D-S theory is prone to paradoxes in high-conflict scenarios of multimodal fusion.In response,this paper proposes a result-level multimodal fusion method that integrates a confidence estimation network with an improved D-S theory.The method consist of two key components.First,a confidence estimation network reframes the classification problem in target detection as a confidence estimation task,providing confidence scores for the detection results of individual unimodal networks.Second,a fusion method with improved D-S theory uses confidence scores and class information to construct evidence,achieving final fusion of detection data from different modalities.Evaluation experiments on the KITTI dataset show that the proposed fusion method improves mAP by up to 6.64%compared to image-based detection and up to 15.43%compared to point cloud-based detection.In the comparison of fusion methods,the proposed fusion method achieves an mAP improvement 0.81%higher than the classical D-S fusion.It effectively reduces classification conflicts and addresses the limitations of the classical D-S theory.

关 键 词：置信估计 环境感知 多模态目标检测 D-S证据理论 结果级融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]