动态场景下基于光流的语义RGBD-SLAM算法

Semantic RGBD-SLAM in dynamic scene based on optical flow

为解决传统的同时定位与建图算法在复杂动态环境下容易受到动态目标干扰而导致定位精度差和建图错误的问题,提出了一种动态场景下基于光流的语义RGBD-SLAM算法。首先,通过优化的二维相邻帧透视矫正方法,对当前帧进行透视矫正以补偿相机运动;然后,将矫正后的图像输入RAFT-S网络中,在获得低分辨率的稠密光流场后提取动态目标的掩码,并根据上一帧掩码中动态目标的位置和速度信息,对当前掩码中的动态区域进行跟踪和优化,从而提取动态目标在每一帧中的精确区域;最后,分离静态和动态特征点,通过最小化静态特征点的重投影误差,得到优化后的相机位姿,并结合轻量级语义分割网络BiSeNetv2提供的语义信息和相机提供的深度信息,构建无人的静态语义八叉树地图。公开数据集TUM上的测试结果表明,本文算法的绝对轨迹误差相对于ORB-SLAM2减少了90%以上,并能获取精确的动态区域掩码以及准确的语义地图,验证了该算法在复杂动态场景中具有良好的定位精度和鲁棒性。

To address the problems of poor positioning accuracy and mapping error in the traditional simultaneous localization and mapping algorithms under the complex dynamic environments with dynamic objects,a semantic RGBD-SLAM algorithm in dynamic scenes is proposed,which is based on the optical flow.Firstly,the camera ego-motion is compensated by the optimized 2D perspective correction method based on adjacent frames.Secondly,by feeding the compensated perspective images into the RIFT-S network,the low-resolution dense optical flow field is obtained for extracting the current mask of the dynamic region.The dynamic regions in the current mask are tracked and optimized by using the position and velocity of the dynamic regions in previous mask.The accurate dynamic regions in each frame can be extracted.Finally,the static and dynamic features are separated,and the optimized camera pose is obtained by minimizing the reprojection error of the static feature points.The static semantic octree map without people is established by the depth data from camera and semantic information produced by the lightweight semantic segmentation network BiSeNetv2.Compared with ORB-SLAM2,the test results on the public data set of TUM indicate that the absolute trajectory error of the proposed algorithm is reduced by more than 90%,and the accurate masks of dynamic regions and an accurate semantic map also can be obtained.Results show that the proposed algorithm has a good positioning accuracy and robustness under complex dynamic scenes.