Research On Simultaneous Localization And Mapping Algorithm Based On RGB-D Camera

In recent years,the improvement of computer processing power has greatly promoted the development of computer vision.With the rapid development of autonomous driving,augmented reality and robotics research,one of its key technologies: vision-based simultaneous localization and mapping has received extensive attention from researchers at home and abroad.The technology intelligently analyzes and calculates the image data acquired by the visual sensor,estimates the pose of the visual sensor,and realizes the positioning of the sensor itself while constructing the incremental map.Compared with the traditional laser sensor-based simultaneous localization and mapping methods,the visual sensor is economical,and the amount of information acquired is large,and the scope of application is wide.Therefore,there exists important the oretical significance and application value to make a depth research of vision-based simultaneous localization and mapping.This paper firstly introduces the research status of vision-based simultaneous localization and mapping algorithms,analyzes the advantages and disadvantages of various typical algorithms,focuses on the simultaneous localization and mapping algorithms based on RGB-D camera,and studies five iterative closest point algorithms.And realize the 3D map construction under the static rigid scene;further study the surface construction method based on the deformation field.The main research work of the thesis is as follows:(1)Aiming at the problem of accurate camera position estimation simultaneous localization and mapping algorithm based on RGB-D camera,this paper conducts in-depth research on the most widely used iterative closest point algorithm in point cloud registration,and compares four iterative closest point algorithms.Based on this,this paper studies the pointto-plane ICP algorithm based on CUDA parallel acceleration,and then selects the best registration and the shortest time algorithm from the five ICP improved algorithms,which lays a solid foundation for fast and accurate camera pose estimation.(2)For the existing real-time 3D reconstruction algorithm KinectFusion can not solve the problem of large scene map reconstruction,this paper proposes a three-dimensional map construction method of static rigid scene.The method reconstructs the map over a long distance by moving the fixed 3D model space in KinectFusion.At the same time,the loop detection and back-end optimization module is added,so that it can detect the closed loop and perform global optimization,reduce the accumulated error,and ensure the consistency of the constructed map.(3)Aiming at the problem that the dynamic target in the scene is difficult to deal with based on the simultaneous localization and mapping algorithm,this paper proposes a surface construction method based on the deformation field.The method estimates the non-rigid transformation by using the adjacent two-frame dynamic target point cloud,and then uses the sparse node to represent and build the deformation field.The deformation field is continuously expanded and updated,so that the new point cloud data is merged into the surface model of the dynamic target.This method is used to model the dynamic target,so that its influence on the construction of the static scene map can be avoided.