Indoor Multidimensional Reconstruction Based on Maximal Cliques

Three-dimensional reconstruction is an essential skill for robots to achieve complex operation tasks, including moving and grasping. Applying deep learning models to obtain stereoscopic scene information, accompanied by algorithms such as target detection and semantic segmentation to obtain finer labels of things, is the dominant paradigm for robots. However, large-scale point cloud registration and pixel-level labeling are usually time-consuming. Here, a novel two-branch network architecture based on PointNet features is designed. Its feature-sharing mechanism enables point cloud registration and semantic extraction to be carried out simultaneously, which is convenient for fast reconstruction of indoor environments. Moreover, it uses graph space instead of Euclidean space to map point cloud features to obtain better relationship matching. Through extensive experimentation, our method demonstrates a significant reduction in processing time, taking approximately one-tenth of the time required by the original method without a decline in accuracy. This efficiency enhancement enables the successful execution of downstream tasks such as positioning and navigation.