【作者】 李翔；

【导师】 姜欣；

【作者基本信息】 哈尔滨工业大学 ， 机械工程， 2021， 硕士

【摘要】 建筑行业作为劳动密集型产业,在人口老龄化和劳动力成本上升的背景下,正面临劳动力短缺的问题。机器人技术的发展使得许多传统行业摆脱了对劳动力的依赖走向自动化,但是其在建筑领域的运用仍然有限。建筑场景的定位不同于一般的室内场景,首先,建筑场景是复杂和高动态的,传感器噪声会对定位造成干扰;其次,为了按照建筑图纸进行施工,定位信息需要与建筑图纸保持坐标系一致;最后,建筑任务需要的定位精度更高,需要对机器人进行高精度的位置跟踪。针对以上特点,设计了一套基于建筑图纸和多传感器融合的定位系统,搭建了传感器平台,并在现实建筑环境中对平台和算法进行了测试。针对建筑环境高动态的特点,设计了基于相机和激光雷达融合的感知方案。提出了图像驱动的点云语义分割的方案,使用图像语义分割网络和数据集对建筑环境进行感知,并通过投影关系对点云进行联合语义分割,解决了建筑场景点云数据集缺失的问题。根据点云语义分割的结果,针对性地过滤动态物体的噪声信息,减少对于定位的影响。针对定位信息与建筑图纸坐标一致的需求,对建筑图纸中的三维模型信息进行采样,得到先验点云地图,并使用点云配准算法进行全局定位,以保证定位与建筑图纸的坐标系统一。在点云配准的流程中,使用基于点云特征描述子的粗配准方法提升配准效率,并使用重叠比例参数评估最终配准效果,同时保证了点云配准算法的效率和准确性。为了提高机器人在位置跟踪中的定位精度,首先将定位问题建模成状态估计问题,使用无迹卡尔曼滤波器对激光雷达里程计和惯性测量单元(Inertial Measurement Unit,IMU)进行融合。通过IMU对机器人的位姿进行预测,并使用激光雷达里程计的观测信息进行更新,提高了机器人的定位精度。搭建了一套多传感器实验平台,完成了多传感器平台的标定。在真实建筑场景中开展了相关实验,证明了相机与激光雷达融合进行语义分割和全局定位方案的可行性,并通过与现有定位算法的对比,证明该系统可以在室内建筑场景实现更高的定位精度。更多还原

【Abstract】 As a labor-intensive industry,the construction industry is facing a labor shortage in the context of an aging population and rising labor costs.The development of robotics technology has allowed many traditional industries to get rid of their dependence on labor and move towards automation,but its application in the construction field is still limited.The localization in the construction scene is different from the general indoor scene.First,the construction scene is complex and highly dynamic,and sensor noise will interfere with the localization.Second,to carry out construction following the drawings,the localization needs to keep the coordinate system consistent with the construction drawings.Finally,high-precision position tracking of the robot is the precondition to perform construction tasks.For the above situation,a localization system based on construction drawings and multi-sensor fusion is designed and tested in real construction environment.Aiming at the high dynamic characteristics in construction scene,a sensing system based on the fusion of camera and Li DAR is designed.Besides,a scheme of imagedriven point cloud semantic segmentation is proposed.The image semantic segmentation network and dataset are used to perceive the construction environment,and the point cloud is jointly segmented through the projection relationship,which solves the problem of missing point cloud dataset in construction scenes.According to the results of point cloud semantic segmentation,the noise information of dynamic objects is filtered to reduce the impact on localization.In response to the requirement that the localization information is consistent with the coordinates of the construction drawings,the 3D model of the construction drawings is sampled to obtain a priori point cloud map.Then,the point cloud registration algorithm is used for global localization to ensure the consistency of coordinate system.In the point cloud registration process,the coarse registration method based on point cloud feature descriptors is used to improve the registration efficiency,and the overlap fraction parameter is used to evaluate the final registration result.In order to improve the localization accuracy of the robot in position tracking,the localization problem is first modeled as a state estimation problem.The unscented Kalman filter is used to fuse the Li DAR odometry and the inertial measurement unit(IMU).The pose of the robot is predicted through the IMU,and the observation information of the Li DAR odometry is used to correct the pose,which improves the localization accuracy of the robot.A multi-sensor platform is built and calibrated.Relevant experiments are carried out in real construction scenes,which proved the feasibility of the camera and Li DAR fusion for semantic segmentation and global localization.Through the comparison with existing localization algorithms,it is proved that the system can achieve higher localization accuracy in indoor construction scenes.更多还原