基于物体关系描述的单目语义SLAM方法  

Monocular Semantic SLAM Method Based on Object Relation Description

作　　者：林士琪 王纪凯 裴浩渊 赵皓 陈宗海[1] Lin Shiqi;Wang Jikai;Pei Haoyuan;Zhao Hao;Chen Zonghai(Department of Automation,University of Science and Technology of China,Hefei 230027,China)

机构地区：[1]中国科学技术大学自动化系,安徽合肥230027

出　　处：《系统仿真学报》2022年第2期278-284,共7页Journal of System Simulation

基　　金：国家自然科学基金(91848111,61703387)。

摘　　要：外界环境的语义感知和自身位置的准确估计是移动机器人自主导航和作业的关键。提出了一种基于单目相机的语义SLAM(simultaneous localization and mapping)方法,在轨迹估计的同时完成三维目标检测。提取物体自身语义、尺寸、颜色分布及其邻域拓扑结构等多元信息作为描述子,实现帧间物体的准确关联。在后端对相机位姿、地图点和物体路标进行联合优化,并自适应调整代价函数中各误差项的权重系数,以提高各状态变量的估计精度和鲁棒性。实验结果表明,所提出的算法在地图构建方面具有较高的精度。Semantic information perception of the external environment and accurate positioning are the keys to autonomous navigation and operation of mobile robots. This paper proposes a method of semantic simultaneous localization and mapping(SLAM) based on a monocular camera. The system completes three-dimensional(3 D) object detection while estimating the trajectory. We model the 3 D objects with cuboids. Then, the semantic meanings, color distribution, size and neighborhood topology of the objects are extracted as descriptors for the accurate matching of objects between different frames. The camera pose, map points and object landmarks are optimized jointly in the backend of the system. The weight coefficient of each error term in the cost function is autonomously adjusted to improve the estimation accuracy and robustness of each state variable of the system. The experimental results show that the proposed method has high accuracy in map construction.

关 键 词：图像分割 三维目标检测 拓扑地图 图匹配 语义SLAM 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]