Research On SLAM Algorithm Based On RGB-D Camera

The demand of the intelligent robot industry drives the rapid development of robotics.The role of navigation in robots is very important.Visual navigation has become the mainstream development direction of robot navigation technology.In order to accomplish related tasks,the robot must first locate its own location and explore and perceive the surrounding environment,then plan the optimal path,and complete the relevant tasks after reaching the exact location.The ordinary monocular camera cannot measure the depth of the object.The binocular camera is expensive and has a large amount of calculation.The Microsoft released Kinect camera can acquire the depth of the object and is widely used because of its low price.Typical SLAM systems are divided into front-end,back-end,loopclosing,and mapping.The front-end is affected by feature extraction and image matching.The back-end optimization determines the global consistency of the map and robot trajectory.The linear optimization method is the mainstream method of back-end optimization in the current stage.Loopclosing can correct the robot trajectory,map and reduce accumulated errors.The specific research content of this paper is as follows:First,the background and significance of the SLAM algorithm based on RGB-D camera are analyzed.The research status of the subject is reviewed,the problems and challenges of the current SLAM based on RGB-D camera are explained,and the main research and framework of the paper are given.Second,introduce the internal composition and working principle of the Kinect camera,introduce the most commonly used camera models and the camera's internal and external parameters,analyzes the sources of camera distortion,and mathematically constructs the camera's distortion.The calibration principle of the Kinect camera is studied.The drift is used to correct the parallax.The reticle is programmed to use the checkerboard to calibrate the Kinect camera to obtain the internal and external parameters and distortion parameters of the RGB camera and the IR camera,and the corrected calibration result is performed,analysis.Third,front-end of the SLAM algorithm is mainly studied.The front-end framework of typical SLAM is introduced.Three common typical features and their principles are studied.Feature extraction with low efficiency,GPU is used to accelerate feature extraction,and grid feature matching method can be used to reduce false matching of images.3D-2D space projection to find the motion of continuous frames,introduced the PnP algorithm to solve the camera's six degrees of freedom pose.The experimental part implements GPU extraction of SIFT features and compares the time difference between CPU and GPU extraction.The experiment compares the extraction results of three features,and compares them with scales,rotations,and fuzzy scenes using different matching methods.At the same time,an experiment is performed on the motion estimation of the front-end continuous frames,and the result of motion estimation is used to splice point cloud maps.And analyze the error.Fourth,the back-end theory of SLAM algorithm is mainly studied,and relevant experimental verification is performed.Two kinds of state estimation methods based on extended Kalman filter and nonlinear optimization are introduced.The solving principle of the mainstream nonlinear optimization method is mainly studied.The reasonable setting of objective function covariance matrix in nonlinear optimization can improve the accuracy of the optimization.The principle of loopclosing based on the appearance of the image is studied,and various forms of map construction and their uses are briefly introduced.The experimental part adopts the TUM data set to realize visual front-end and back-end optimization experiment based on improved covariance.The results obtained by the two methods are used to map the point cloud and analyze the error.In the laboratory scenario,the Amigo robot is equipped with a PC and a Kinect camera for field experiments.Finally,I analyze and summarize all the work done in this paper,and put forward the direction of the next stage and the problems that need to be solved.