动态环境下基于延迟语义的RGB-D SLAM算法  

RGB-D SLAM Algorithm Based on Delayed Semantic Information in Dynamic Environment

作　　者：王浩[1] 周申超 方宝富[1] WANG Hao;ZHOU Shenchao;FANG Baofu(School of Computer Science and Information Engineering,Hefei University of Technology,Hefei 230601)

机构地区：[1]合肥工业大学计算机与信息学院,合肥230601

出　　处：《模式识别与人工智能》2023年第10期953-966,共14页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61872327);安徽省自然科学基金项目(No.2308085MF203);安徽高校协同创新项目(No.GXXT-2022-055);民航飞行技术与飞行安全重点实验室重点项目(No.FZ2022ZZ02);民航飞行技术与飞行安全重点实验室开放基金项目(No.FZ2022KF09)资助。

摘　　要：目前,将分割网络与SLAM(Simultaneous Localization and Mapping)结合已成为解决视觉SLAM不能应用于动态环境的主流方案之一,但是SLAM系统受到分割网络处理速度的限制,无法保证实时运行.为此,文中提出基于延迟语义的RGB-D SLAM算法.首先,并行运行跟踪线程与分割线程,为了得到最新的延迟语义信息,采取跨帧分割的策略处理图像,跟踪线程根据延迟语义信息实时生成当前帧的语义信息.然后,结合成功跟踪计数(STC)与极线约束,筛选当前帧动态点的集合,并确定环境中先验动态物体的真实运动状态.若确定该物体在移动,继续将物体区域细分为矩形网格,以网格为最小单位剔除动态特征点.最后,利用静态特征点追踪相机位姿并构建环境地图.在TUM RGB-D动态场景数据集及真实场景上的实验表明文中算法在大部分数据集上表现较优,由此验证算法的有效性.Visual simultaneous localization and mapping(SLAM)cannot be applied to dynamic environment.The mainstream solution is to combine segmentation network and SLAM.However,real-time operation of SLAM systems is not guaranteed due to the processing speed constraints of segmentation network.Therefore,a RGB-D SLAM algorithm based on delayed semantic information in dynamic environment is proposed.Firstly,tracking and segmentation threads run in parallel.To obtain the latest delayed semantic information,a cross-frame segmentation strategy is employed for image processing,and real-time semantic information for the current frame is generated by the tracking thread according to the delaysemantic information.Then,the dynamic point set of the current frame is selected and the real motion state of the prior dynamic object in the environment is determined by combining successful tracking count and epipolar constraints.When the object is determined as moving,the object area is further subdivided into rectangular grids,and dynamic feature points are removed with the grid as the minimum unit.Finally,the camera pose is tracked by static feature points and an environment map is constructed.Experiments on TUM RGB-D dynamic scene dataset and real scenes show that the proposed algorithm performs well and its effectiveness is verified.

关 键 词：同时定位与地图构建(SLAM) 动态环境 实例分割网络 延迟语义信息 网格聚类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]