动态场景下仓储机器人的视觉定位与建图  

Vision SLAM of storage robot in dynamic scene

作　　者：徐兴 刘琼[1] 黄开坤 XU Xing;LIU Qiong;HUANG Kaikun(School of Mechanical Engineering,University of South China,Hengyang,Hunan 421200,China)

机构地区：[1]南华大学机械工程学院,湖南衡阳421200

出　　处：《导航定位学报》2023年第6期110-118,共9页Journal of Navigation and Positioning

基　　金：国家重点研发计划项目(2023YFC3010900);南华大学教学改革项目(2019ZD-XJG09)。

摘　　要：针对目前大多数室内仓储机器人的视觉定位与建图算法(SLAM)是假设机器人处在静态的环境,但是当场景中出现移动的物体时,机器人自身定位的准确性和稳定性易受到巨大影响,而其他室内定位技术比如超宽带、蓝牙等必须在无线信号覆盖的条件下工作的问题,提出一种面向室内仓储机器人在动态场景下的视觉SLAM算法:在基于旋转不变特征点的定位与建图改进算法(ORB-SLAM2)基础上,用几何对应神经网络(GCNv2)来替换基于图像金字塔的特征点提取算法;添加目标检测(YOLOv4)的语义线程,并且使用光流法来追踪特征点;然后通过运动一致性检测来识别图像中潜在的动态物体;最后剔除动态特征点后进行位姿估计。实验结果表明,在高度动态的室内场景下,提出的算法相对于ORB-SLAM2算法的绝对轨迹误差可减小95.56%~98.21%,能够有效解决ORB-SLAM2在动态场景下定位不准确的问题。Aiming at the problems that the current majority of visual simultaneous localization and mapping(SLAM)algorithms utilized by indoor storage robots are predicated upon the assumption that the robot be operating within a static environment,however,the presence of mobile objects within the visual field can greatly impair the accuracy and stability of the robot's inherent positioning capabilities,while other indoor positioning technologies such as ultra wide band(UWB)and bluetooth must work under the condition of wireless signal coverage,the paper proposed a vision-based SLAM algorithm tailored specifically to indoor storage robots operating within dynamic scenes:the geometric correspondence network version 2(GCNv2)was leveraged to supplant the traditional oriented fast and rotated brief(ORB)feature point extraction algorithm based on image pyramids based on the ORB-SLAM2 framework;and the semantic threads of target detection technique you only look once version 4(YOLOv4)were added in tandem with optical flow to track feature points;then the potential dynamic objects were identified in the image through motion consistency detection;finally,after the removal of dynamic feature points,the position and attitude estimation was performed.Experimental result showed that in highly dynamic indoor scenes,the absolute trajectory error of the proposed algorithm,compared to the ORB-SLAM2 algorithm,would be reduced by 95.56%~98.21%,which indicates that the proposed method could effectively solve the problem of inaccurate positioning of ORB-SLAM2 in dynamic scenarios.

关 键 词：仓储机器人 动态场景 几何对应神经网络(GCNv2) 目标检测 基于旋转不变特征点的定位与建图改进算法(ORB-SLAM2) 

分 类 号：P228[天文地球—大地测量学与测量工程]