Optimization Of RGB-D SLAM System Based On Feature Points

With the development of computer vision,human-machine interaction and perception of 3D world are needed.However,the 2D images obtained by cameras as the main image acquisition equipment are ambiguous,which cannot support the computer vision to enter a more advanced 3D stage.Meanwhile,more and more applications such as virtual reality(VR),service robot navigation,augmented reality(AR)and so on require the support of the underlying 3D dense point cloud map.In order to restore the 3D coordinates of the scene,the inverse projection of the image needs to be reconstructed.SFM,the most commonly used technique in traditional 3D reconstruction,needs to consumes a huge amount of memory and computing resources in the reconstruction,and the recovery is a 3D sparse point cloud,and dense reconstruction is needed to recover the dense structure.Visual SLAM technology can rapid reconstruction of 3D point cloud.Visual SLAM collects the data of the surrounding environment by camera for 3D reconstruction,while RGB-D camera is a kind of sensor that can collect the RGB information and depth information of the environment at the same time.Compared with other sensors,it is more conducive to the recovery of the scene's dense structureIn this paper,considering the current RGB-D SLAM system problems such as poor realtime performance,failure of initialization,inaccurate pose estimation of camera,and "drift" of 3D model,etc.A robust RGB-D SLAM system based on feature points is proposed to optimize the above problems.The RGB-D camera is used as a sensor to collect data,and the initial pose of the camera is calculated and optimized by extracting and matching the feature points of the image,finally,according to the correspondence between the point cloud data and the camera coordinate system,the point clouds of all depth image restoration are converted to the same coordinate system,so as to achieve the reconstruction of 3D dense point clouds.The main contents of this paper are as follows:(1)In this paper,aiming at the problems of poor matching accuracy and slow matching speed of the existing feature matching algorithms,the feature detection algorithm is improved and the extracted feature points are processed by using the feature point homogenization strategy based on quadtree.Combining with the Bag-of-Words,the scope of feature matching is limited in the feature matching stage,so that the accuracy of matching is improved.Meanwhile,the feature matching speed is ensured by the k-fork tree.Experimental results show that the proposed method is faster and more accurate than traditional algorithm.(2)On the basis of feature matching,a method of automatic system initialization is proposed.In order to deal with the single texture in the collected initial image,two initialization models were constructed.In the system initialization stage,the score of the model is calculated,the initialization model is selected according to the score,the pose of the camera is estimated by using the twodimensional feature matching relation,and the 3D space points corresponding to the feature points are recovered by triangulation method.(2)On the basis of feature matching,a method of automatic system initialization is proposed.In order to deal with the situation that the single texture in the collected initial image and many pixels with a value of 0 in the depth image,two initialization models are constructed.In the system initialization stage,the score of the model is calculated,the initialization model is selected based on the score,the pose of the camera is estimated by using the two-dimensional feature matching relation,and the 3D space points corresponding to the feature points are recovered by triangulation method.(3)Aiming at the problem of inaccurate pose estimation and "drift" of reconstruction results.When the system is initialized successfully,the method of RANSAC + P3P+ nonlinear optimization was used to estimate the initial value of camera pose from frame to frame,so as to provide a better initial value for local and global Bundle Adjustment of key frames.In the subsequent optimization,in order to reduce the optimized data dimension,a common view is used to filter the image and select the key frames in the image sequence.And on the basis of the key frame,the system's repositioning function is built to ensure that the system can still recover camera pose by key frame in the case of tracking loss.At the same time,loop detection is used to add loop constraints to the system to reduce the cumulative error.Experimental results in TUM dataset show that the proposed algorithm is better than RGB-D SLAM v2 in pose estimation accuracy,reconstruction results and reconstruction time.Finally,the 3D dense reconstruction is conducted in the indoor scene with a large number of single textures,and the results show that the proposed optimization method has better robustness and real-time performance.