基于Deeplab V3+的机器人语义SLAM算法研究

Research on robot semantic SLAM algorithm based on Deeplab V3+

针对机器人在同步定位与地图构建(SLAM)系统中受几何场景信息计算力和带宽负载的限制,对ORB-SLAM2框架进行改进,提出语义跟踪和语义建图线程,语义跟踪线程通过Deeplab V3+对图像语义分割,同时提取该图像特征点,进行移动一致性检查来剔除动态噪声点,结合一致性检查后的特征点和分割后的图像信息来二次检查动态点,随后位姿估计,而语义建图线程主要完成语义八叉树地图的构建。在TUM RGB-D数据集上进行了广泛实验,在walking系列数据中的旋转漂移误差达到1.19m、平移漂移误差达到0.046m,满足实时性要求,所提方法有效提高了SLAM的精度和鲁棒性。

In order to solve the limitation of the computing power and bandwidth load of geometric scene information in the Simultaneous Localization And Mapping(SLAM)system of the robot,the ORB-SLAM2 framework is improved,and the semantic tracking and semantic mapping thread are proposed.The semantic tracking thread uses Deeplab V3+to do image semantic segmentation,extract the image feature points at the same time,and perform the movement consistency check to remove the dynamic noise points.After that,the feature points after the consistency check and the segmented image information are combined to check the dynamic points twice,and then estimate the pose.The semantic mapping thread mainly completes the construction of the semantic octree map.Extensive experiments have been carried out on the TUM RGB-D data set.In the walking series data,the rotation drift error reaches 1.19m and the translation drift error reaches 0.046m,which meets the real-time requirements.The proposed method effectively improves the accuracy and robustness of SLAM.