| With the development of technology,emerging applications such as indoor robots,virtual reality,and augmented reality have gradually entered people's lives.Visual Simultaneous Localization and Mapping(V-SLAM)technology plays a vital role in this field.It is a research hotspot in the field of robotics and computer vision.The visual odometer is the most critical part of the visual SLAM system.At present,in the mainstream visual SLAM system,the visual odometer part simultaneously estimates the rotational motion and the translational motion.Without using other SLAM techniques(loopback detection,global relocation,etc.),the path trajectory has a large drift error,and it is difficult to output other valid visual information.Therefore,there are high requirements for the subsequent optimization work.In order to solve these technical problems,this paper carried out related research work in the indoor Manhattan scene.(1)The rotation motion is estimated by the density distribution of the space surface normal vector and the re-projection error function is defined by the image feature points to estimate the translational motion.The accuracy of the pose estimation is finally improved.(2)For the limitations in the basic algorithm,the improvement is based on it,so that the system can still estimate the rotational motion when only one type of plane and parallel lines in the plane are seen.In the process of solving the rotational motion,the linear geometric information in the space is added.In the translational motion solving process,the feature points without depth information are added.The accuracy of the pose estimation is finally improved.(3)In the visual odometer part combined with the network model to initially identify the two-dimensional topology in the indoor Manhattan scene,an advanced twodimensional topological line fitting and node optimization algorithm is proposed and implemented to output effective visual information. |
