融合动态目标跟踪的视觉SLAM算法  

Visual SLAM Algorithm of Integrating Dynamic Target Tracking

作　　者：白克强[1,2] 朱亚兰 杨秀清 向勇 邓子犇 姜官武 BAI Keqiang;ZHU Yalan;YANG Xiuqing;XIANG Yong;DENG Ziben;JIANG Guanwu(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China;Engineering Research Center of Integration and Application of Digital Learning Technology,Ministry of Education,Beijing 100039,China;Civil Aviation Logistics Technology Company Limited,Chengdu 610199,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621010 [2]数字化学习技术集成与应用教育部工程研究中心,北京100039 [3]民航成都物流技术有限公司,四川成都610199

出　　处：《信息与控制》2024年第5期574-584,共11页Information and Control

基　　金：四川省科技计划重点研发项目(2023YFG0061);数字化学习技术集成与应用教育部工程研究中心创新基金项目(1331003);绵阳市应用技术研究与开发项目(2021ZYZF1004)

摘　　要：为提高机器人在动态场景下的SLAM(simultaneous localization and mapping)精度,同时实现对动态目标的跟踪,提出了一种融合动态目标的视觉SLAM方法。首先,通过预处理模块获取RGB(red, green, blue)图像的光流向量、实例分割结果及深度图信息;其次,迭代求解相机位姿、地图点和动态目标位置的初始值;最后,通过一种改进的因子图优化方法对3种状态变量的初始值进行联合优化。在KITTI数据集上的测试实验结果表明,该算法实现了融合动态目标跟踪的视觉SLAM功能,同时有效地提高了动态目标的跟踪精度及动态场景下的SLAM精度,总体效果优于VDO-SLAM(visual dynamic object-aware SLAM)。To improve the accuracy of simultaneous localization and mapping(SLAM)for robots in dynamic environments and enable dynamic target tracking,we propose a visual SLAM algorithm that integrates dynamic target tracking.First,we obtain the optical flow vector,instance segmentation results,and depth map information of red-green-blue(RGB)images through a preprocessing module.Second,we iteratively solve for initial values of camera pose,map points,and dynamic target positions.Finally,we jointly optimize these initial values using an improved factor graph optimization method.Experiments conducted on the KITTI dataset demonstrate that our algorithm successfully combines visual SLAM with dynamic target tracking,significantly improving the accuracy of dynamic target tracking and SLAM in dynamic scenes.The overall performance surpasses that of visual dynamic object-aware SLAM(VDO-SLAM).

关 键 词：同步定位与地图构建 动态场景 目标跟踪 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]