运动模糊场景下结合感知增强与特征约束的SLAM算法
SLAM algorithm combining perceptual enhancement and feature constraints in dynamic blurry scenarios
陈孟元 1程浩 1李鹏飞 1郭行荣1
作者信息
- 1. 安徽工程大学 电气工程学院,芜湖 241000;高端装备先进感知与智能控制教育部重点实验室,芜湖 241000
- 折叠
摘要
针对传统同时定位与地图构建(SLAM)算法在运动模糊场景下难以准确识别物体和无法有效判断物体的实际运动状态等问题,提出一种结合感知增强与特征约束的视觉SLAM算法.设计了一种单阶段去模糊识别的网络模型,通过引入模糊区域关注和增强识别机制识别环境中的物体并获取语义信息,提高了系统在复杂环境下的感知能力.为进一步减少动态物体对系统定位精度的影响,融合语义信息和特征约束,提出了一种基于全局条件随机场的运动判断模型.在公开的 TUM 数据集上进行验证,结果表明所提算法有效提高了运动模糊环境下物体识别的精度和准确率,平均绝对轨迹误差与ORB-SLAM3、DS-SLAM、Dyna-SLAM算法相比分别减少了 96.3%、51.4%以及 10.2%,表现出良好的构图能力.
Abstract
Aiming at the problems that traditional simultaneous localization and mapping(SLAM)are difficult to accurately recognize objects and unable to effectively judge the actual motion state of objects in dynamic blurry scenarios,a SLAM algorithm combining perceptual enhancement and feature constraints is proposed.The algorithm designs a network model for single-stage deblurring recognition,which recognizes objects in the environment and acquires semantic information by introducing the fuzzy region attention and enhancement recognition mechanism,improving the perceptual ability of the system in complex environment.To further reduce the influence of dynamic objects on the localization accuracy of the SLAM system,a motion judgment model based on global conditional random field is proposed by fusing semantic information and feature constraints.The validation results on the public TUM dataset show that the proposed algorithm effectively improves the precision and accuracy of object recognition in dynamic blurry scenarios,and the average absolute trajectory error is reduced by 96.3%,51.4%and 10.2%,respectively,compared with ORB-SLAM3,DS-SLAM and Dyna-SLAM algorithms,which exhibits a good composition capability.
关键词
同时定位与地图构建/语义信息/特征约束/运动判断Key words
simultaneous localization and mapping/semantic information/characteristic constraint/motion judgment引用本文复制引用
基金项目
安徽省重点研究与开发计划项目(202304a05020073)
安徽省高校协同创新项目(GXXT-2021-050)
安徽省高校杰出青年科研项目(2022AH020065)
安徽省学术和技术带头人后备人选科研活动经费择优资助(2022H292)
出版年
2024
NETL
NSTL
万方数据