The Method of Edge Optimization for Indoor Scene Reconstruction and Virtual-Real Occlusion Based on VSLAM

In augmented reality environment, the fusion effect of virtual and real objects is often affected by virtual-real occlusion. To improve the occlusion effect, this article proposes a method for dense reconstruction and segmentation of three-dimensional objects in indoor scenes based on VSLAM. The method first uses YOLOv5s and ORB-SLAM2 to detect and remove dynamic feature points in the environment, to build an accurate point cloud map using only static feature points. Then, the OPTICS clustering algorithm is used to constrain the voxel edges and perform mesh segmentation. Finally, the proposed method predicts the reconstruction of the segmented point cloud by combining the shape prior algorithm, thereby making the edges of the segmented objects more accurate. The proposed method was tested on multiple datasets, and experiments dynamic feature points removal and virtual-real occlusion were performed. The results indicate that, compared to traditional ORB-SLAM2 in dynamic scenarios, the positioning accuracy of the camera has improved by 92.62%, and the reconstruction accuracy of point clouds has improved by 35.00%. This suggests that the method can accurately locate the occlusion edges of virtual and real objects and then segment them, while maintaining a shaped reconstruction result, thus making the occlusion effects between virtual and real objects more realistic and natural.