Research On Fusing Depth Information Direct SLAM Based On Photometric Calibration

With the rapid development of science and technology,more and more mobile robots appear in people's production and life.A number of researchers are going on SLAM(Simultaneous Localization and Mapping,SLAM)that is one of the key technologies for mobile robots to autonomous navigation.Visual SLAM can better obtain the texture information of the scene,compared with traditional laser SLAM has more advantages in map construction,intelligence and cost.There are many algorithms for normal vision SLAM,in which direct method vision SLAM has some disadvantages such as: poor adaptability and low robustness,at this stage.For this reason,this article calibrates innovatively the Kinect camera by photometry and designs a direct method visual SLAM algorithm that fuses depth information for the mobile machines actual operating scenes with features of few feature points,lighting transformations and large scenes.It effectively improves the robustness and positioning accuracy of the algorithm when illumination changes and promotes the development of visual SLAM technology for mobile robot autonomous navigation.Firstly,the Kinect 2.0 camera was calibrated by internal parameters and photometry.Photometric calibration of Kinect 2.0 camera is an innovation.The optical principle of the camera is discussed,and then the optical vignetting model is analyzed.The polynomial model and vignetting compensation model were used to model the camera response function and vignetting function,respectively.Finally,the two function were coupled to obtain a photometric calibration model.The imaging principle of Kinect 2.0 camera is studied,and the camera model,coordinate system transformation,Kinect 2.0 structure and parameters are described in detail.Using the Zhang calibration method to calibrate the Kinect 2.0 camera and obtain its internal parameters.Secondly,the traditional direct method vision SLAM fused the depth information to solve the camera pose and simulation.The depth information is added to the traditional direct method which only uses the photometric information,and the error function is constructed together,after comparing the direct method and the feature point method.The nonlinear optimization model is used to solve the least square problem.What is more,the camera pose is obtained.Thirdly,simulating the back-end optimization algorithm and map construction algorithm.This paper uses the Pose Graph map optimization algorithm as the back-end for the characteristics that the direct method does not extract feature points,after comparing three mainstream visual SLAM back-end optimization algorithms.A dense point cloud map was constructed using the point cloud mosaic algorithm,for the needs of the project.In order to facilitate storage and navigation,the point cloud map is filtered and converted into an octree map for easy navigation and storage.TUM public dataset was used to verify the effects of pose estimation algorithm,back-end optimization algorithm and map reconstruction algorithm.Finally,experiments are designed for photometric calibration and visual SLAM.In a laboratory scene,a standard grayscale board was used to perform a photometric calibration of Kinect 2.0 to obtain its photometric function.Then,using SVM algorithm to conduct image recognition experiments,and the results show that the photometrically calibrated image is recognized more effectively.The visual SLAM algorithm was verified on the ball picking robot platform,which is independently developed by the laboratory.The lighting conditions were changed and experiments were performed,in the laboratory scene,The experimental results show that the improved direct visual SLAM algorithm in this paper has better robustness and improved accuracy by 0.015 m.