面向动态环境的双目/惯性SLAM系统  

Stereo/Inertial SLAM system for dynamic environment

作　　者：张乐添 赵冬青[1] 贾晓雪 杨朝永 赖路广 郭文卓 ZHANG Letian;ZHAO Dongqing;JIA xiaoxue;YANG Chaoyong;LAI Luguang;GUO Wenzhuo(School of Geodesy and Geomatics,Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]信息工程大学地理空间信息学院,郑州450001

出　　处：《导航定位学报》2022年第6期144-150,共7页Journal of Navigation and Positioning

基　　金：国际自然科学基金项目(41774037,42104033)。

摘　　要：基于静态场景假设的即时定位与建图(SLAM)系统仍存在诸多问题:传统视觉SLAM中的关键帧并不包含语义信息,在实际场景中总包含大量动态点,不仅会影响系统精度,甚至可能出现跟踪失败。针对以上问题,提出一种语义SLAM思路。首先,提出利用掩膜区域卷积神经网络(Mask R-CNN)给出的先验语义信息进行筛选,摒弃动态点的干扰,将这项功能作为新的线程加入尺度不变特征变换即时定位与建图(ORB-SLAM3)。其次,结合ORB-SLAM3的双目/惯性里程计部分,基于惯导预积分协方差设置自适应阈值,判断先验动态点的真实动态性。最后,利用实测数据对算法进行验证。实验结果表明,3个数据集序列绝对轨迹误差分别降低56.3%、71.5%、19.4%。因此,该系统能够有效判别特征点的真实动态性,屏蔽动态点的干扰,从而提高定位精度。There are still many problems with the static scenario based on simultaneous localization and mapping(SLAM):the keyframes in traditional visual SLAM do not contain semantic information, but a large number of dynamic points in the actual scenario that not only affect system accuracy, but also easily track failure. This paper proposes a semantic slam system. First, the interference of the dynamic point is removed with prior semantic information given by mask region convolutional neural network(Mask R-CNN), the function join oriented fast and rotated brief simultaneous localization and mapping(ORB-SLAM3) as a new thread.Secondly, combining the stereo/inertial odometry, the adaptive threshold is set based on the inertial measurement unit(IMU) preintegration covariance, and the real dynamics of the transcendental dynamic point is determined. Finally, the experimental results show that the absolute trajectory errors of the three datasets are reduced by 56.3%, 71.5% and 19.4%,respectively. Therefore, the system effectively distinguish the real dynamics of feature points, shield the interference of d ynamic points, and improve the positioning accuracy.

关 键 词：动态环境 即时定位与建图 预积分 语义分割 神经网络 

分 类 号：P228[天文地球—大地测量学与测量工程]