南京邮电大学学报（自然科学版）2026，Vol.46Issue(3)：51-61,11.DOI:10.14132/j.cnki.1673-5439.2026.03.006
AGRF-Net:基于边界感知与自适应门控残差融合的RGB-D语义分割网络
AGRF-Net:adaptive gated residual fusion network with edge awareness for RGB-D semantic segmentation
摘要
Abstract
The reliability of environment perception systems is the foundation upon which autonomous driving technology highly depends.However,in complex road scenes,single-modal perception struggles to simultaneously address the detection challenges posed by drastic illumination changes and the pres-ence of both positive and negative obstacles.To improve the accuracy and robustness of cross-modal fea-ture fusion,this paper designs an adaptive gated residual fusion network(AGRF-Net).In this network,auxiliary edge information and multi-dimensional adaptive gating mechanisms are utilized to optimize the complementarity and integration process of cross-modal features.First,an edge-aware input enhancement mechanism is introduced to extract auxiliary edge features,and the network's perception capability of the fine geometric contours of obstacles is strengthened through a multi-scale gated edge fusion module.Second,a depth multi-level feature enhancement module is utilized to perform internal refinement and contextual enhancement of depth features,thereby suppressing the inherent noise of the depth modality and supplement geometric details.Third,a cross-modal residual gating network is constructed as the core fusion module,which filters effective depth information through multi-scale contextual reliability gating and achieves high-order feature integration by combining a cross-modal attention mechanism.Finally,ex-perimental results validate the feasibility and accuracy of the proposed strategy in handling the collabora-tive detection of both positive and negative obstacles in complex road environments.关键词
自动驾驶/RGB-D语义分割/正负障碍物/自适应门控/特征融合Key words
autonomous driving/RGB-D semantic segmentation/positive and negative obstacles/adap-tive gating/feature fusion分类
信息技术与安全科学引用本文复制引用
徐鹤,张恩俊,谭萍..AGRF-Net:基于边界感知与自适应门控残差融合的RGB-D语义分割网络[J].南京邮电大学学报（自然科学版）,2026,46(3):51-61,11.基金项目
江苏省前沿技术研发计划(BF2025617)资助项目 （BF2025617）
