面向室内动态场景的视觉同时定位与地图构建语义八叉树地图构建方法

Visual Simultaneous Localization and Mapping Method of Semantic Octree Map Toward Indoor Dynamic Scenes

针对传统视觉同时定位与地图构建(vSLAM)系统在动态场景中无法有效去除运动物体及缺少可用于更高层应用的语义地图等问题,提出了一种可有效去除动态物体并构建表征室内静态环境的语义八叉树地图的vSLAM系统方法。首先,使用FastSCNN作为语义分割网络提取图像的语义信息,同时,利用金字塔光流法对特征点进行跟踪匹配。然后,使用步进随机抽样一致算法(Multistage RANSAC)通过多次执行不同尺度的RANSAC流程对特征点进行步进采样,再利用对极几何约束并结合FastSCNN提取的语义信息进行视觉里程计动态特征点剔除。最后,通过体素滤波降低点云冗余后构建纯静态环境的语义八叉树地图。实验结果表明:所提方法在公用数据集TUM RGBD的8个RGBD高动态序列中测试的相机相对位移误差、相对旋转误差和全局轨迹误差相较于ORBSLAM2系统有94%以上的提升,全局轨迹误差仅为0.1 m;相较于同类DSSLAM系统,动点剔除总耗时有21%的缩减。建图性能方面,经体素滤波后构建的语义点云地图与语义八叉树地图分别占据9.6 MB、685 kB的存储空间,相较于17 MB的原始点云,语义八叉树地图仅占用其4%的存储空间并因含有语义可用于更高层次的智能交互任务。

Aiming at the problems that traditional visual simultaneous localization and mapping(vSLAM)systems cannot remove moving objects in dynamic scenes effectively and lack semantic maps for highlevel interactive applications,a vSLAM system scheme was proposed.The scheme can remove moving objects effectively and build semantic octree maps representing indoor static environments.First,FastSCNN was used as a semantic segmentation network to extract semantic information from images.Meanwhile,a pyramid optical flow method was used to track and match feature points.Then,for step sampling of the feature points,a stepping random sampling consistent algorithm(Multistage RANSAC)was used to perform the RANSAC process on different scales several times.Later,the epipolar geometry constraint and semantic information extracted from the FastSCNN were combined to remove the dynamic feature points of the visual odometer.Finally,the semantic octree map representing the static indoor environment was built by the point cloud after using voxel filtering to reduce redundancy.Experimental results show that the performance indicators of a camera,including relative displacement,relative rotation,and global trajectory errors in the 8 RGBD high dynamic sequence of common datasets TUM RGBD,are improved by more than 94%compared with the ORBSLAM2 system,and the global trajectory error is only 0.1 m.Compared with a similar DSSLAM system,the total time for eliminating a moving point is reduced by 21%.After voxel filtering,the semantic point cloud and octree maps occupy 9.6 MB and 685 kB storage space,respectively,in terms of map construction performance.Compared with the original point cloud of 17 MB,the semantic octree map occupies only 4%of the storage space;therefore,it could be used for highlevel intelligent interactive applications due to its semantics.