融合语义分割与光流网络的动态鲁棒SLAM算法OA
Integrating semantic segmentation and optical flow for robust SLAM in dynamic scenes
针对动态场景下传统视觉同步定位与建图(SLAM)系统因特征关联错误导致定位漂移的问题,本文在提出了一种融合语义分割与光流估计的增强型语义SLAM系统――SF-SLAM.该系统构建了双流感知机制以鲁棒地解耦动态干扰:一方面利用轻量级语义分割网络(RTMDet-Ins-tiny)精准剔除人、车等先验动态物体;另一方面引入光流估计网络(LiteFlowNet2)并结合自适应OTSU阈值算法,有效捕获并抑制手持物体、移动家具等非先验动态特征.为进一步提升建图的静止性与可靠性,本文设计了一种基于KD树加速的欧氏聚类后处理策略,通过跨帧几何一致性校验机制,精准识别并移除语义与光流模块遗漏的残余动态噪点,最终生成几何精确且语义清晰的静态八叉树地图.这一地图形式去除了动态干扰,特别适用于机器人高层导航与路径规划任务.在TUM RGB-D基准数据集及包含强动态干扰(如人为移动广告牌)的真实场景中进行了广泛实验.结果表明,相比于现有主流方法,该算法在各类动态等级下均实现了定位精度的显著提升,并在复杂现实环境中表现出优异的鲁棒性与实时性能.
Addressing the vulnerability of traditional visual simultaneous localization and mapping(SLAM)systems to dynamic occlusions and feature mismatches,this paper proposes an enhanced semantic SLAM framework built upon ORB-SLAM3.A dual-stream perception mechanism is introduced to robustly decouple dynamic interference.Specif-ically,a lightweight semantic segmentation network(RTMDet-Ins-tiny)is employed to generate pixel-wise masks for a priori dynamic categories(e.g.,humans,cars),while an optical flow network(LiteFlowNet2)combined with adaptive OT-SU thresholding is utilized to detect and suppress non-prior dynamic motion(e.g.,carried items,shifted furniture).To ensure map consistency,a post-processing pipeline utilizing k-dimensional(KD)-tree accelerated Euclidean clustering is developed.This mechanism identifies and eliminates residual dynamic artifacts missed by the front-end modules through cross-frame geometric verification,yielding a highly clean static semantic octree map suitable for autonomous navigation.Comprehensive evaluations on the TUM RGB-D benchmark and challenging real-world scenarios featuring active interfer-ence demonstrate that the proposed method significantly outperforms state-of-the-art approaches in positioning accuracy and robustness across various dynamic levels while maintaining competitive real-time performance.
刘尊;杨骏;许湘珺;王利;罗琦;黄畅昕;李坚强
深圳大学 人工智能学院,广东 深圳 518060深圳大学 人工智能学院,广东 深圳 518060深圳大学 人工智能学院,广东 深圳 518060深圳大学 人工智能学院,广东 深圳 518060中广核集团有限公司 安全壳检测实验室,广东 深圳 518124深圳大学 人工智能学院,广东 深圳 518060深圳大学 人工智能学院,广东 深圳 518060
语义SLAM光流估计语义分割动态场景移动机器人
semantic SLAMoptical flowsemantic segmentationdynamic environmentmobile robot
《控制理论与应用》 2026 (9)
1890-1900,11
国家自然科学基金杰出青年科学基金项目(62325307),国家自然科学基金项目(62527809,62373258),深圳重点产业研发计划项目(ZDCY20250901102300001),深圳市科技项目(KJZD20230923115215032,JCYJ20240813141628038)资助. Supported by the National Science Fund for Distinguished Young Scholars(62325307),the National Natural Science Foundation of China(62527809,62373258),the Shenzhen Key Industry Research and Development Program(ZDCY20250901102300001)and the Shenzhen Science and Technology Program(KJZD20230923115215032,JCYJ20240813141628038).
评论