Research On Key Issues And Methods Of Depth Camera V-SLAM In Indoor Environment

The continuous economic and social progress has made the indoor robot industry closely related to human life develop rapidly.As a core problem in the field of robotics,SLAM(Simultaneous Localization And Mapping)technology has attracted wide attention from scholars.The depth camera sensor used in the indoor environment,with the advantages of both price and function,makes the depth camera V-SLAM(Visual SLAM)technical problem become a research hotspot.In order to solve some problems of the existing technology of the depth camera V-SLAM,the key problems and methods of depth camera data acquisition and depth image repair,key frame extraction and loop detection,V-SLAM dense mapping and optimization are studied in this paper.Applying it to actual indoor scenes,the experimental analysis and verification of the technical methods studied in this paper are carried out.The main work of this article is as follows:(1)Aiming at the problem of depth camera sensor model and data acquisition,this paper designs an image acquisition method suitable for Kinect v1.First,the pinhole camera model of the depth camera Kinect v1 is described,the coordinate transformation and mathematical modeling of the camera space points are discussed,and the camera calibration parameters are derived,and the Kinect v1 camera is calibrated using Zhang's calibration method(“checkerboard-method”).Finally,the image representation in computer graphics was explained in detail.On the Windows platform,Kinect development kit and MATLAB image processing tools were used to complete the V-SLAM data acquisition and conversion of indoor environment.(2)Aiming at the noise and removal of V-SLAM input data,this paper proposes a set of depth image enhancement method flows.First,the causes of the noise in the depth image are analyzed and summarized.The FMM(Fast Marching Method)algorithm is improved,and the inverse threshold binary method of the image is used to generate the repair mask.Then,the hole in the depth map is repaired with neighborhood estimate method.Then,an image smoothing method based on median filtering is used to further enhance the edge noise of the depth map.Finally,the quality of the repaired image is evaluated to verify that the method in this paper can improve the quality of the depth map by 17.1%.(3)Aiming at the accuracy and efficiency of V-SLAM motion estimation,this paper uses the ORB feature operator for feature extraction and registration of color images.On the key issues of V-SLAM back-end module,a key frame filtering method based on feature correlation is designed to select key frames of V-SLAM images to improve the system's operating efficiency and quality of construction.Finally,a loop detection method based on the ORB feature bag of words is studied,and a comparison experiment with the bag of words model DBoW library is performed.The experiments prove that the self-trained bag of words model has good reliability and robustness for loop detection.(4)Aiming at the dense mapping of V-SLAM in an indoor environment,a dense mapping method of depth camera V-SLAM is designed.First,the key frame selection mechanism of this paper is applied to construct a three-dimensional dense map through key frame point cloud mosaic.Secondly,a point cloud filtering optimization method is adopted to remove noise and redundancy in the map,and finally convert the point cloud map into an octree(3D OctoMap)map.Finally,the well-known TUM dataset and the actual indoor environment scene are used to experiment the proposed mapping method.The results verify that the proposed method can remove about 10%of the redundant point clouds.At the same time,the frame rate of drawing is 26.1%higher than RGB-D SLAM v2,and explain that the method in this paper can clearly and efficiently achieve the construction of dense maps.