Research And Realization Of 3D SLAM Based On RGB-D Camera Indoor

With the popularization of smart applications such as mobile robots in people's daily lives,the level of intelligence in mobile robots has also continued to increase.The fast and accurate positioning is one of the key technologies in the current research of intelligent mobile robots.A mobile robot can use a GPS system to accurately locate itself outdoors,but in an indoor or unknown environment,it can only use its own sensor to perform a synchronous positioning to create a map in real time which is called SLAM(Simultaneous Location And Mapping).In present,the RGB-D depth camera can directly obtain pixel depth information,simplifying the complicated depth value calculation process.Therefore,research on indoor three-dimensional SLAM based on RGB-D is a hotspot at home and abroad.This paper puts forward a new type of dense SLAM system with ORB keyframes based on the RGB-D depth camera which optimizes the modules of rigid motion transformation,visual odometry,key frame selection strategy,pose diagram construction,scene recognition,loopback detection,back-end graph optimization and map creation.This topic comes from the joint project of the Shenzhen Intelligent Robot Research Center of the Chinese Security Research Institute and the Chinese University of Hong Kong--"3D Mapping with an RGB-D Camera Project",the main work of this paper is as follows:1.For the uncertainty of the depth information obtained by the depth camera at non-effective distance,a pre-processing method based on bilateral filtering is proposed to reduce the influence of the camera model.2.A fusion motion estimation method is proposed.In the tracking thread of ORB-SLAM2,the direct method based on pixels is used for motion tracking,which avoids the time-consuming defect of feature extraction from the original feature point and optimizes the selection strategy of key frames.3.A dynamic depth-information binding RANSAC-ICP method is proposed to dynamically use the weights of the effective depth information to participate in constructing the RANSAC optimal model,removing outliers that don't meet in the sample data points and finally the RANSAC Coarse matching results is used for ICP fine-match motion estimation to obtain the transformation matrix.4.Based on the sparse map constructed by ORB-SLAM2,a dense map is created and finally comes up a dense point cloud map with consistent environment.Meanwhile,an octree map which is easy to store and maintainis created and provides the necessary prerequisites for subsequent navigation and path planning.The experimental result show that the proposed method has the following advantages: good tracking performance,accurate motion estimation,and globally consistent trajectory map;the system can operate in real-time environment since the time of visual odometer is greatly reduced;good robustness;Loopback can be correctly detected to generate an easy-to-maintain and updatable octree map.