Research On Indoor Map Construction Based On RGB-D SLAM Algorithm

At present,indoor maps are not as rich as outdoor maps.Most indoor maps are based on two-dimensional CAD plane maps,and their visual effects are poor.Due to the lack of systematic guiding principles and unified standards for the study of indoor maps,the construction methods of indoor maps and the degree of information acquisition have not been paid enough attention,so far it is difficult to assist people in their daily life.Conventional three-dimensional laser scanner can complete the construction of indoor three-dimensional model,but the laser sensor is huge and expensive,and it is not suitable for indoor environment.Traditional close-range photogrammetry technology can achieve environmental data acquisition by repeating stand,but the operation is cumbersome and the data processing is time-consuming and laborious.In this paper,the Simultaneous Localization and Mapping(SLAM)technology in the field of intelligent robots is studied in depth.SLAM technology is applied in the field of Surveying and mapping to solve the problem of real-time construction of indoor maps.In this paper,we will study the application of location in building inexpensive equipment that can quickly detect environmental changes and develop corresponding applications.The real-time construction of indoor maps based on consumer-level visual sensors is studied in detail.Using Microsoft’s RGB-D camera Kinect as an environmental sensing sensor,the SLAM technology is studied theoretically and analyzed experimentally.This paper mainly completes the following work:(1)The pinhole camera model of Kinect,a commonly used RGB-D camera,is described.The composition of Kinect camera and the principle of depth detection are briefly introduced.Focusing on the source of camera distortion and its mathematical modeling,this paper lays a theoretical foundation for subsequent camera calibration.Using the checkerboard to calibrate the Kinect camera to obtain the camera’s internal parameters and distortion parameters,the data acquisition experiment was completed using the calibrated camera.At the same time,the color map and depth map aligned after calibration are obtained.(2)The theoretical basis of the RGB-D SLAM scheme is studied,including: Relative motion estimation based on feature point method in front-end visual odometer,mainly for feature detection operator performance analysis,feature extraction and initial matching based on ORB algorithm.At the same time,the traditional 3D camera-based visual odometer implementation method is introduced.The principle and problems of the implementation of ICP algorithm based on 3D point cloud are clarified.The back-end graph optimization and the word bag-based loopback detection process are described in detail.(3)In the aspect of eliminating mismatches,this paper proposes a feature point matching optimization method based on RANSAC algorithm.At the same time,the pose of the camera is estimated in real time by combining P3 P with ICP.An implementation method of visual odometer based on SBA algorithm is proposed to reduce the cumulative error of the system and ensure the accuracy and reliability of the whole system.The algorithm of visual SLAM based on nonlinear optimization is studied,and a reasonable key frame selection mechanismis proposed to reduce the cumulative error of visual odometer.A comparative experiment is carried out on three common feature detection algorithms of SIFT,SURF and ORB.Their performances are compared in terms of real-time performance and high efficiency.The RGB-D SLAM system software and hardware experimental platform was built and different feature detection algorithms were applied to the RGB-D SLAM scheme for specific comparison experiments.The RANSAC algorithm is used to eliminate the mismatched feature points and combined with the ICP algorithm to realize the visual odometer estimation based on feature points.The RGB-D SLAM experiment using Turtlebot robot in indoor scene completed the real-time construction of three-dimensional point cloud map of indoor environment.The accuracy of point cloud map is evaluated by coordinate transformation.The experimental results show that the point accuracy of point cloud map is 0.0398 m,and the dimension accuracy of feature in point cloud map is 0.0135 m.After octree processing,it can serve the later stage of indoor autonomous navigation.