Research On Key Technologies Of SLAM Based On RGB-D Images

Nowadays,artificial intelligence,virtual reality and machine learning are growing rapidly.Computer-vision-based robotics has been greatly developed.As the core issue in this field,simultaneous localization and mapping has been aroused wide concern of scholars from all over the world.With the development of the sensitive element and the vision sensor technology,there has been numbers of wellused,reliable and commercial depth sensors,which greatly reduce the difficulty of the research on SLAM based on RGB-D images.A new round of SLAM research begins to rise.This paper is guided by computer vision theory,does some research on the key technologies of SLAM based on RGB-D images.The main work of this research is as follows:1.The depth sensor Kinect V2 is firstly introduced in this paper,and we fix the bug of depth date visualized program which Microsoft provided.Then we take the camera imaging model as the starting point,introduce the camera calibration principle and the solving method.After that,the calibration experiment of the Kinect camera is carried out,the Kinect images is registered and displayed with 3D point cloud.2.In order to ensure the accuracy and reliability of the RGB-D information,the error source of depth image acquired by Kinect is analyzed.Then we research the denoising algorithm of depth image according to the filtering denoising technique in digital image processing.Through the quantitative analysis of experimental results,we find that bilateral filtering algorithm can effectively reduce the noise of depth image.3.The RGB-D SLAM visual odometry based on feature points is realized.If the SIFT feature is extracted by CPU,it takes a long time and can not satisfy the real-time requirement of SLAM.Therefore,a SIFT feature extraction algorithm based on GPU acceleration is used.The experiment shows that GPU-SIFT we designed is 6-9 times faster than CPU-SIFT in a feature extracting period.On this basis of SIFT features,the method of estimating the 3D pose of the object from the 2D point map is used to estimate the camera pose.4.The back-end optimization algorithm based on extended Kalman filter is introduced.This method was popularly used in SLAM several years ago.In this paper,we adopt a graph optimization algorithm which can effectively solve the problem of cumulative error in large scale map creation.By constructing the pose graph based on RGB-D image and adding the loop closure,we achieved a full SLAM process.Then,the proposed system is evaluated using standard RGB-D dataset.The result shows that the error between the estimated trajectory and the real trajectory can be controlled within 0.1m under the condition of sufficient image acquisition,which proves that globally consistent environment map can be built by the SLAM system.