RGB-D Visual SLAM Algorithm Based On Image Features

Simultaneous localization and mapping is the key algorithm to carry out autonomous control of robots. SLAM algorithm combined with vision sensors is a big breakthrough in this field, which makes robots sense their surroundings like creatures. In recent years, visual SLAM algorithm based on images has achieved increasingly development.6-DOF (degree of freedom) position optimization to 3D visual SLAM technology of different kinds of robots, such as home robots, underwater robots, etc., is important for research and economic development.In this paper, the RGB-D visual SLAM system uses Kinect as a visual sensor. Kinect can provide color images and the depth images of the indoor suroundings at the same time. This article involving RGB-D visual SLAM algorithm based on image features mainly solves the following problems:1) Feature extraction and matching of color images. Specific procedures are utilized to extract features from color images, to compute feature descriptors and to match features of the key frames. The feature extraction algorithms using in this paper are SIFT (Scale Invariant Feature Transform), SURF (Speeded Up Robust Features) and the ORB (ORiented Brief). Then the number of features, running time and accuracy rate by using these three algorithms are compared to find the best feature extraction method to complete our algorithm.2) Calculating 3D coordinate of the features. Firstly, finding depth information of features of color images in the corresponding depth images. Secondly, we calculate 3D world coordinates of these features according to relative algorithm. Finally, extrinsic parameter matrixes of the camera can be obtained by utilizing RANSAC or RGBD-ICP, which is decomposed to relative displacement and relative euler angles.3) Global optimization and closed-loop detection. This article uses EKF-SLAM to optimize global path and global 3D map, and uses closed-loop detection to reduce global error. Experiment processes at a rate of 20HZ, satisfying the real-time requirements.4) Experimental results and analysis. Combined with the existing data-sets, the 3D trajectory of the robot will be obtained. Then two methods are proposed to validate this SLAM algorithm:relative pose error and absolute trajectory error. The smaller the error is, the closer the SLAM trajectory is to the real trajectory. Then 3D maps obtained using SLAM algorithm of this paper and using SFM (Structure From Motion) algorithm are compared visually.Through the above process, we get trajectory and 3D map of the environment of the robots which is globally optimized. The experiment results show that the error between SLAM trajectory and true trajectory is small, and this algorithm satisfies the real-time requirements.