Research On Visual SLAM Technology Based On RGB-D Camera

With the development of computer technology,intelligent mobile robots are used more and more frequently in various fields.Simultaneous localization and mapping(SLAM)is the key to realize the autonomy of mobile robots when performing various complex operations.SLAM using vision sensor is a hot research direction at present,among which RGB-D camera can capture depth image and color image at the same time provides a new development direction for visual SLAM.In this paper,the key problems of RGB-D SLAM,such as low matching accuracy of feature points,poor back-end optimization and environmental adaptability,are analyzed and studied.The specific work is as follows.(1)The principle of data acquisition of RGB-D camera is analyzed.The calibration principle of depth camera and the alignment of depth image with color image are introduced.Based on Intel D435 depth camera,the parameter calibration and the alignment of depth image with color image are realized,the accuracy of depth measurement is analyzed at the same time.Three different types of visual SLAM are implemented based on D435.Experiments show that RGB-D SLAM can determine scale information compared with monocular SLAM and is more efficient than binocular SLAM.(2)At the front end of SLAM,there are many mismatching problems when ORB feature points are matched in visual odometer.In this paper,the combination of bidirectional matching and RANSAC algorithm is used to improve the matching accuracy.The matching correct rate reaches 95.2%,and compared to the one-way matching with the RANSAC algorithm increased by 5%.According to the results of the matching algorithm in this paper,the visual odometer is realized by EPnP algorithm.The root mean square error of the result is 0.085 m,which is 29% higher than that of the one-way matching plus RANSAC algorithm.(3)At the back end of SLAM,the pose estimation and 3D construction are optimized respectively.Firstly,for the problem of poor optimization of traditional Extended Kalman Filter.The BA algorithm is adopted to optimize the graph of pose and spatial point.Then,for the problem of cumulative error of visual odometer,construct the pose based on key frame and add loopback detection for global optimization,so as to further eliminate the accumulated errors.The root mean square error before optimization is 0.085 m,and the root mean square error after back-end optimization is 0.033 m.The accuracy increased by 41%..Finally,the original three-dimensional point cloud map is filtered.(4)The RGB-D SLAM is implemented while the accuracy of the RGB-D SLAM system is evaluated by standard TUM data sets,and the engineering implementation in different environments is analyzed and Application.First,the absolute error of the RGB-D SLAM pose is about 4cm which measured by the standard data set.In the actual indoor environment,the camera is positioned and the indoor environment is reconstructed by RGB-D SLAM with a handheld camera.Then,in the actual indoor environment,the RGB-D SLAM is operated by means of a handheld camera and a crawler-based mobile robot,and an indoor three-dimensional map is constructed.Finally,in view of the failure of ordinary visual SLAM in dark environment,only the depth image generated by the depth information of RGB-D camera is used,and the block matching algorithm is used for depth image matching and pose calculation.Experiments show that the depth information can basically realize the pose calculation and preliminary mapping in the dark environment.