| The 3D modeling method using RGB-D sensors can easily obtain depth information,which has great advantages over traditional monocular or binocular methods.However,under the current development of the 3D real-time modeling technology based on the RGB-D method,it also faces some technical difficulties and challenges,such as the balance of computility,real-time performance and accuracy.Usually,algorithms that can meet real-time performance will have a large loss in accuracy.The improvement of realtime accuracy often requires the computing power support of parallel computing devices such as GPUs.Based on this,this paper studies the real-time 3D modeling technology in indoor scenes.In the end,the accuracy of the system was improved while ensuring low computility and real-time operation.The specific research contents include:1.An improved real-time image feature is proposed,which combines the depth information of RGB-D sensor on the basis of ORB feature,and calculates descriptors on the depth map and grayscale image respectively.This feature uses the image pyramid on the depth image and the grayscale image to achieve the scale invariance of the feature,and uses the grayscale centroid method to achieve the rotation invariance.In order to make the distribution of feature points in the image more reasonable,a feature point homogenization algorithm is designed.2.Aiming at the improved features,a feature matching algorithm is designed for it,which performs feature matching according to the comprehensive distance between the gray scale descriptor and the depth descriptor.In addition,in order to better eliminate false matches,a false match elimination algorithm based on texture information and spatial position information of feature points is proposed.It can be seen from the experimental results that the features in this thesis have better performance than ORB features.3.In the sparse mapping algorithm,in addition to the method of back-projection of feature points in RGB-D mapping,a triangulation method based on 2D matching points in monocular mapping is also used,which effectively increases the number of sparse map points.In the dense mapping algorithm,methods such as error point elimination,statistical filtering and voxel filtering are introduced,and finally a better point cloud map is obtained.4.A 3D real-time modeling system has been built,which can run in real time with the support of ordinary CPU computing power.Use the Intel RealSense D435 sensor to scan the indoor scene in real time,and calculate the sensor pose trajectory and 3D point cloud map visually in real time.Using multiple RGB-D datasets of different difficulty on the TUM dataset,it is compared with the pose trajectory results of ORB SLAM2 and ORB SLAM3.Finally,experiments were conducted in real indoor scenes,and the modeling accuracy was measured using the MeshLab tool.The experimental results show that the 3D real-time modeling algorithm in this thesis can effectively improve the accuracy of pose trajectory estimation and modeling under the condition of low computing power and real-time performance.. |
