动态场景结合稀疏场景流和加权特征的视觉SLAM方法  

Visual SLAM method combining sparse scene flow and weighted features in dynamic environment

作　　者：闫河 王旭 雷秋霞 YAN He;WANG Xu;LEI Qiuxia(School of Artificial Intelligence,Chongqing University of Technology,Chongqing 401135,China)

机构地区：[1]重庆理工大学两江人工智能学院,重庆401135

出　　处：《中国惯性技术学报》2024年第9期891-897,共7页Journal of Chinese Inertial Technology

基　　金：国家自然科学基金面上项目(61173184);国家重点研发计划“智能机器人”重点专项(2018YFB1308602)。

摘　　要：针对视觉同时定位和建图(SLAM)系统大多基于静态环境假设、在动态场景位姿估计不准确的问题,提出了一种动态场景结合稀疏场景流和加权特征的视觉SLAM方法。首先,引入实例分割网络SparseInst获取环境中的语义信息,识别潜在运动物体;其次,通过计算特征点的稀疏场景流及其马氏距离获得特征点在相机坐标系下的相对运动,利用卡方检验方法完成对动态特征点的检测和剔除;然后,对静态特征点分配权重,设计加权的光束平差法优化目标函数以解决部分特征点运动状态模糊问题,提高视觉SLAM在动态场景下的定位精度。在公共数据集上的对比实验表明,所提方法相较于ORB-SLAM2、YOLO-SLAM、SG-SLAM,其绝对轨迹均方根误差在TUM RGB-D和Bonn数据集上分别平均降低约94.69%、27.55%、5.27%和93.43%、38.30%、26.88%。Addressing the challenge of visual simultaneous localization and mapping(SLAM)systems predominantly relying on static environment assumptions and exhibiting inaccurate pose estimation in dynamic scenes,a visual SLAM method combining sparse scene flow and weighted features in dynamic environment is proposed.Firstly,the SparseInst instance segmentation network is introduced to obtain semantic information in the environment and identify potential movable objects.Secondly,the relative motion of feature points in the camera coordinate system is obtained by calculating the sparse scene flow of feature points and their Mahalanobis distance,and the detection and removal of dynamic feature points are achieved by using the Chi-square test method.Then,static feature points are assigned weights,and a weighted objective function of the bundle adjustment optimization is designed to address the issue of ambiguous motion states of some feature points,thereby enhancing the localization accuracy of visual SLAM in dynamic scenes.Comparative experiments on the public datasets show that compared with ORB-SLAM2,YOLO-SLAM and SG-SLAM,the absolute trajectory root mean square error of the proposed method is reduced by an average of about 94.69%,27.55%,5.27%and 93.43%,38.30%,26.88%on the TUM RGB-D and Bonn datasets,respectively.

关 键 词：视觉SLAM 动态场景 稀疏场景流 加权特征 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]