基于多模态数据的注意特征融合姿态估计网络  

Pose estimation network based on attention feature fusion of multimodal data

作　　者：赵云涛 邓新辉 ZHAO Yuntao;DENG Xinhui(College of Information Science and Engineering,Wuhan University of Science and Technology.,Wuhan 430081,China)

机构地区：[1]武汉科技大学信息科学与工程学院,湖北武汉430081

出　　处：《液晶与显示》2025年第4期598-606,共9页Chinese Journal of Liquid Crystals and Displays

基　　金：湖北省工程研究中心开放课题(No.IDICP-KF-2024-08)。

摘　　要：兼顾准确性与适用性的6D姿态估计一直是研究热点和难点。本文提出了一种基于多模态数据的注意特征融合的6D姿态估计网络。首先,引入了更深层次的挤压激励模块结构,通过调整各通道的权重来增强依赖性扩大感受野,提升处理RGB图像特征的效果。进一步,针对多模态数据,在特征融合阶段部署迭代注意特征融合模块,通过多次迭代融合操作解决全局特征融合中的尺度不一致问题,能够更准确地捕捉和整合多模态数据,显著改善了姿态回归的效果。最后,为了量化评估模型在复杂环境下的鲁棒性和适用性,引入了不可见百分比指标,该指标能够评估模型在处理部分遮挡或复杂背景时的性能。通过在公开数据集上的姿态预测实验,验证了改进后模型不仅能够在验证数据集上实现了准确的预测姿态,而且相较于Densefusion模型,本文提出的算法模型在复杂环境下更具适用性。6D pose estimation that balances accuracy and applicability has been a hot and difficult research topic.To this end,a 6D pose estimation network based on attentional feature fusion of multimodal data is proposed.Firstly,a deeper structure of squeeze and excitation module is introduced to enhance the dependency to expand the receptive field by adjusting the weights of each channel to improve the effect of processing RGB image features.Further,for multimodal data,an iterative attention feature fusion module is deployed in the feature fusion stage,which solves the scale inconsistency problem in global feature fusion through multiple iterative fusion operations,and is able to capture and integrate multimodal data more accurately,which significantly improves the effect of attitude regression.Finally,in order to quantitatively assess the robustness and applicability of the model in complex environments,an invisibility percentage metric is introduced,which is capable of assessing the performance of the model when dealing with partially occluded or complex backgrounds.Through the pose prediction experiments on the public dataset,it is verified that the improved model is not only able to achieve accurate predicted poses on the validation dataset,but also the algorithmic model proposed in this paper is more applicable in complex environments compared to the densefusion model.

关 键 词：6D姿态估计 多模态数据 注意特征融合 不可见百分比 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]