一种动态场景下语义分割优化的ORB_SLAM2

为提高ORB_SLAM2在动态环境下位姿估计的准确性，提出使用语义分割剔除分布在人上的移动特征点进而提高位姿准确性的方法.该方法在对输入图像提取ORB特征点的同时，对图像进行语义分割，获得人在图像中的像素点位置，再将这些分布在人上面的特征点进行剔除，并使用剔除后相对稳定的特征点进行位姿估计.利用改进方法在TUM公共数据集上进行测试，结果表明，改进后的系统能够明显降低动态环境下位姿估计的绝对误差和相对漂移，证明了该方法较原始的ORB_SLAM2系统能够明显提高动态环境下位姿估计的准确性.
In order to improve the accuracy of ORB_SLAM2 in dynamic scene, this paper proposed a method of using semantic segmentation to eliminate the mobile feature points distributed on the human body to improve the accuracy of pose. This method extracted ORB feature points from the input image, semantically segmented the image to obtain the position of the pixels in the image, and then eliminated these feature points distributed on the top of the human body, and estimated the pose using the relatively stable feature points after eliminating. The improved method was tested on the TUM data set, and results show that it can reduce absolute error and relative drift under dynamic environment, which proves that the method is more accurate in pose estimation than the traditional in the dynamic scene.