一种椭球模型表示的对象级动态语义SLAM方法  

Object-level dynamic semantic SLAM method based on ellipsoidal model

作　　者：白素琴 诸皓伟[2] 吕宗磊 王成根[3] 史金龙 BAI Suqin;ZHU Haowei;LYU Zonglei;WANG Chenggen;SHI Jinlong(Key Laboratory of Smart Airport Theory and System,Civil Aviation University of China,Tianjin 300300,China;School of Electrical and Information Engineering,Jiangsu University of Science and Technology,Suzhou 215600,China;School of Computer Science,Jiangsu University of Science and Technology,Zhenjiang 212003,China)

机构地区：[1]中国民航大学民航智慧机场理论与系统重点实验室,天津300300 [2]江苏科技大学电气与信息工程学院,苏州215600 [3]江苏科技大学计算机学院,镇江212003

出　　处：《中国惯性技术学报》2025年第1期46-54,共9页Journal of Chinese Inertial Technology

基　　金：国家自然科学基金委员会面上项目(51875270);中国民航大学民航智慧机场理论与系统重点实验室开放基金资助(SATS202207)。

摘　　要：为了提高动态场景同步定位与建图(SLAM)的相机位姿精度和鲁棒性,提出一种高精度对象级SLAM方法。首先,将检测到的实例对象表示为3D椭球模型,构建对象地图;接着,寻找实例对象和地图中对象之间的最佳匹配关系;然后,通过运动检测找到场景中的动态对象,在地图中追踪对象的运动轨迹,并不断更新其对应的3D椭球模型,以确保对象地图的准确性;最后,采用静态对象和内部3D点联合优化的方式,在跟踪丢失后重新定位相机。在TUM和BONN数据集上的实验结果表明:所提方法具有更高的相机位姿精度,位姿误差仅为OA-SLAM算法误差的12.5%、ReFusion算法的16.7%、ACEFusion算法的33.3%。重定位实验结果表明:所提的相机重定位策略有效地解决了动态场景中相机丢失的问题,提高了系统的鲁棒性。代码开源在https://github.com/wawcg/23Object-SLAM。To improve the accuracy and robustness of camera pose estimation in dynamic scene simultaneous localization and mapping(SLAM),a high-precision object-level SLAM method is proposed.Firstly,detected instance objects are represented as 3D ellipsoidal models,and an object map is constructed.Next,the best matching relationships between instance objects and objects in the map are found.Then,dynamic objects in the scene are identified through motion detection,their trajectories are tracked in the map,and their corresponding 3D ellipsoidal models are continuously updated to ensure the accuracy of the object map.Finally,a camera relocation method for dynamic scenes,which combines the optimization of static objects and internal 3D points,is employed to relocate the camera after tracking loss.The experimental results on the TUM and BONN datasets show that the proposed method has higher camera pose accuracy,and the pose error is only 12.5%of the OA-SLAM algorithm,16.7%of the ReFusion algorithm,and 33.3%of the ACEFusion algorithm.Relocation experiments indicate that the proposed camera relocation strategy effectively addresses the issue of camera tracking loss in dynamic scenes,enhancing the system's robustness.The code is available at https://github.com/wawcg/23Object-SLAM.

关 键 词：动态场景 同步定位与建图 数据关联 对象追踪 椭球模型 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置] TH74[自动化与计算机技术—控制科学与工程]