Research On The RGB-D SLAM System Based On Point And Line Features In The Indoor Structured Environment

Simultaneous localization and mapping(SLAM)is the basis of autonomous navigation and obstacle avoidance for mobile robots.Based on the SLAM tech-nology,the robot can use its own sensors to perceive the information of the un-known surrounding environment,and to locate its position in the surroundings and build a map to describe the environment in real time.The SLAM of using the camera as its own sensor is called Visual SLAM.Compared with the lidar sensor,the camera has the advantages of inexpensive and the image contains rich information.Therefore,Visual SLAM has become a research hotspot in the field of robot and computer vision in recent years.Compared with monocular and stereo cameras,RGB-D cameras can provide both the RGB color image and the corresponding depth image of surroundings,which has been widely concerned in recent years.There are abundant point features in the natural environment.In Visual SLAM,the research and application of point features has been mature.However,the traditional Visual SLAM algorithm based on point features only considers the point features in the environment,and the result is not very ideal in the low texture environment.In the artificial structured environment,such as indoor offices and corridors so on,there are abundant line features,which can be used as the supplement of point features.This paper focuses on the issue of Visual SLAM in indoor structured environment,and add line features to SLAM to improve the location accuracy and robustness.Finally,we propose a Visual SLAM system based on both point and line features with RGB-D cameras.The main contents of this paper are as follows:Firstly,the operating principle,the imaging model and the calibration method of RGB-D cameras are introduced.According to the pinhole imaging model and the Zhang Zhengyou calibration method,RGB-D camera is calibrated.And the internal parameters,distortion coefficient and relative pose matrix of the RGB-D camera are obtained.After the calibration of the RGB-D camera,the accuracy of matching between the color image pixel and the depth image pixel is improved obviously.Then,the RGB-D visual odometry based on both point features and line fea-tures is studied.By improving the ORB algorithm to make point features dis-tributing uniformly and adding LSD algorithm to extract line features in the image,the accuracy of frame-to-frame matching is improved.The point features and line features in the image are extracted and matched respectively,and then the error model of combining point features and line features is presented for motion estimation.Finally,the RGB-D SLAM system based on both point features and line fea-tures is studied.A keyframe insertion and elimination method is proposed to maintain the quantity of key frames.A graph optimization model based on both point and line features is presented for back-end optimization,and a visual bag-of-words model based on both point and line features is presented for loop-closure detection.On the basis of the previous front-end visual odometry,the back-end optimization based on both point and line features is added to build a complete visual SLAM system.In this paper,experiments are carried out on the benchmark dataset to verify that our proposed RGB-D SLAM system obtains superior performance on accu-racy and robustness than the point-only system and the line-only system.The experiment in the actual indoor scene shows that our approach can obtain a good performance of localization and mapping in actual run time.