一种融合语义地图与回环检测的视觉SLAM方法  

Loop detection and semantic mapping algorithm fused with semantic information

作　　者：郑冰清 刘启汉 赵凡 张小国[1] 王庆[1] ZHENG Bingqing;LIU Qihan;ZHAO Fan;ZHANG Xiaoguo;WANG Qing(School of Instrument Science and Engineering,Southeast University,Nanjing 210000,China)

机构地区：[1]东南大学仪器科学与工程学院,南京210096

出　　处：《中国惯性技术学报》2020年第5期629-637,共9页Journal of Chinese Inertial Technology

基　　金：国家自然科学基金面上项目(62073078)。

摘　　要：针对机器人导航中动态环境下回环检测易失败且语义地图信息使用不充分的问题,提出了一种融合语义信息的V-SLAM方法,包括语义地图构建和语义回环检测两部分。首先利用YOLOv3深度学习网络对关键帧进行二维图像语义标注;然后基于条件随机场模型将二维图像的语义标注信息与场景分割的聚类信息相融合求解语义地图;语义回环检测算法基于关键帧的语义标注信息,结合运动特征点去除的思路对词典法进行改进,通过去除特定动态物体提高动态场景下的系统定位精度。最后,采用RGB-D TUM数据集以及室内真实场景测试数据验证所提出算法性能。实验表明,所提出的算法在获取更准确语义地图的同时,在室内动态场景下相比使用原算法的系统精度提升了48.1%。Aiming at the problem that loop detection in dynamic environment of robot navigation is easy to fail and insufficient use of semantic map information,a V-SLAM method integrating semantic map construction and semantic loop detection is proposed.In semantic map construction,YOLOv3 deep learning network is used to semantically annotate the two-dimensional image for the key frames,and then based on the conditional random field model,the semantic labeling information is fused with the clustering information of the scene segmentation to solve the semantic map.The semantic loopback detection algorithm based on the semantic annotation of key frames is improved by removing specific dynamic objects.The performance of the proposed algorithm is evaluated using RGB-D TUM SLAM datasets and indoor real scene test data,and experimental results show that the semantic loop algorithm is able to provide better semantic map information,and in indoor dynamic scenarios,improve the accuracy of the SLAM system by 48.1%compared to the original algorithm.

关 键 词：同步定位与建图 动态场景 回环检测 语义地图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]