改进OneFormer分割网络的动态物体实时跟踪SLAM算法  

Dynamic object real-time tracking SLAM algorithm for improved OneFormer segmentation network

作　　者：陈孟元 杨苏朋 许瑞珩 李鹏飞 CHEN Mengyuan;YANG Supeng;XU Ruiheng;LI Pengfei(School of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China;Key Laboratory of Advanced Perception and Intelligent Control for High-end Equipment,Ministry of Education,Wuhu 241000,China)

机构地区：[1]安徽工程大学电气工程学院,芜湖241000 [2]高端装备先进感知与智能控制教育部重点实验室,芜湖241000

出　　处：《中国惯性技术学报》2025年第3期257-266,共10页Journal of Chinese Inertial Technology

基　　金：安徽省重点研究与开发计划项目(202304a05020073);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);安徽省高校杰出青年科研项目(2022AH020065)。

摘　　要：针对传统视觉同时定位与地图构建(SLAM)在动态场景下容易出现特征匹配错误,导致位姿定位与建图精度下降的问题,提出了一种改进OneFormer分割网络的动态物体实时跟踪SLAM算法。首先,在Oneformer主干网络内增加特征增强模块和多注意力模块,增强对潜在动态区域的识别与分割。其次,结合改进后网络获取到的RGB帧中语义信息,使用相机位姿和物体运动偏转角度估计物体运动状态。然后,将识别到的动态物体映射回网络分割掩膜图上形成分层结构,实现对动态物体的精准跟踪。最后,在公开数据集TUM中进行验证。结果表明,与DP-SLAM和DynaSLAM相比,所提算法绝对轨迹误差的均方根误差平均减少了17.70%、19.63%,展现了良好的跟踪精度。Aiming at the problem that traditional visual simultaneous localization and mapping(SLAM) is prone to feature matching errors in dynamic scenes,which leads to a decrease in the accuracy of of pose localization and map construction,a dynamic object tracking SLAM algorithm for improved OneFormer segmentation network is proposed.Firstly,feature enhancement module and multi-attention module are added to Oneformer backbone network to enhance the recognition and segmentation of potential dynamic regions.Secondly,combined with the semantic information in RGB frames acquired by the improved network,the motion state of the object is estimated using the camera pose and the object motion deflection angle.Then,the identified dynamic objects are mapped back to the network segmentation mask map to form a hierarchical structure to achieve accurate tracking of dynamic objects.Finally,validation is performed in the public dataset TUM.The results show that compared with DP-SLAM and DynaSLAM,the root-mean-square error of the absolute trajectory error of the proposed algorithm is reduced by 17.70% and 19.63% on average,showing good tracking accuracy.

关 键 词：同时定位与地图构建 动态场景 深度学习 映射 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]