代军,雷莹莹,赵俊伟,等.基于语义信息与动态补偿的改进型V-SLAM算法[J].河南理工大学学报（自然科学版）,2026,45(6):157-166.

Dai J, Lei Y Y, Zhao J W,et al.Improved V-SLAM algorithm based on semantic information and dynamic compensation[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(6):157-166.

基于语义信息与动态补偿的改进型V-SLAM算法

代军1, 雷莹莹1, 赵俊伟1, 袁兴起1, 王跃功2, 杨茗豪1, 程晓琦3

1.河南理工大学 机械与动力工程学院，河南 焦作  454003；2.平顶山平煤机煤矿机械装备有限公司，河南 平顶山 467000；3.佛山大学 机电工程与自动化学院，广东 佛山  528225

摘要: 目的 针对传统视觉同步定位与建图（V-SLAM）系统在动态环境中定位精度下降的问题，通过融合视觉语义信息与动态补偿技术减少动态物体对定位结果的影响，从而提升无人车在复杂环境中的导航定位精度和鲁棒性。 方法 提出一种基于单目或深度相机的改进型V-SLAM算法。该算法结合相机运动信息与场景语义信息，设计多阶段处理策略。首先，利用几何约束分析筛选静态与动态特征点；其次，引入改进的轻量级深度学习网络识别动态物体区域，为进一步解决短期动态检测遗漏问题，算法结合高斯混合模型（GMM）和卡尔曼滤波（KF）进行动态补偿；最后，通过多级信息融合剔除剩余动态特征点，确保位姿估计仅依赖于静态特征点。  结果 实验结果表明，该改进型算法在多种动态场景下均显著提升了定位性能：在高动态场景中定位精度提高95.5%，低动态场景中提高20.1%，复杂真实场景中提高89.2%。此外，该算法展现出优异的实时性能，能够满足机器人导航系统的实际应用需求。  结论 提出的基于语义信息与动态补偿的V-SLAM算法在动态环境中表现出卓越的定位精度与鲁棒性，同时具有较低的计算复杂度。该算法为动态场景下无人车导航与定位提供了有效的技术解决方案，并为自主移动机器人技术的发展奠定了研究基础。

关键词:视觉同步定位与建图;卡尔曼滤波;语义信息;动态环境;定位精度

doi:10.16186/j.cnki.1673-9787.2024120018

基金项目:国家自然科学基金资助项目（62201151）；河南省自然科学基金资助项目（262300421791）；河南省校企协同创新项目（26AXQXT035）

收稿日期:2024/12/08

修回日期:2025/04/08

出版日期:2026/10/09

Improved V-SLAM algorithm based on semantic information and dynamic compensation

Dai Jun1, Lei Yingying1, Zhao Junwei1, Yuan Xingqi1, Wang Yuegong2, Yang Minghao1, Cheng Xiaoqi3

1.School of Mechanical and Power Engineering， Henan Polytechnic University， Jiaozuo  454003， Henan， China；2.Pingdingshan PMJ Coal Mine Machinery Equipment Co.，Ltd.，Pingdingshan  467000，Henan，China；3.School of Mechatronic Engineering and Automation， Foshan University， Foshan  528225， Guangdong， China

Abstract: Objectives To address the issue of localization accuracy degradation in traditional visual simultaneous localization and mapping （V-SLAM） systems in dynamic environments， this paper integrates visual semantic information with dynamic compensation techniques to reduce the impact of dynamic objects on localization results， thereby improving the navigation and localization accuracy and robustness of unmanned vehicles in complex environments.  Methods An improved V-SLAM algorithm based on monocular or depth cameras is proposed. The algorithm combines camera motion information with scene semantic information and employs a multi-stage processing strategy. First， geometric constraints are used to classify static and dynamic feature points. Second， an improved lightweight deep learning network is introduced to identify dynamic object regions. To address the issue of missed detections in short-term dynamic tracking， the algorithm further incorporates a Gaussian mixture model （GMM） and Kalman filter （KF） for dynamic compensation. Finally， multi-level information fusion is applied to eliminate remaining dynamic feature points， ensuring that pose estimation relies solely on static feature points. Results Experimental results demonstrate that the proposed improved algorithm significantly enhances localization performance in various dynamic scenarios： localization accuracy is improved by 95.5% in high-dynamic scenes and 20.1% in low-dynamic scenes， while an improvement of 89.2% is achieved in complex real-world scenes. In addition， the algorithm exhibits excellent real-time performance， meeting the practical requirements of robotic navigation systems.  Conclusions The proposed V-SLAM algorithm based on semantic information and dynamic compensation achieves outstanding localization accuracy and robustness in dynamic environments while maintaining low computational complexity. This algorithm provides an effective technical solution for unmanned vehicle navigation and localization in dynamic scenes and lays a research foundation for the development of autonomous mobile robot technology.

Key words: visual simultaneous localization and mapping; Kalman filter; semantic information; dynamic environment; localization accuracy