Research on a Dual-Flow Graph Feature Interaction PointCloud Registration Method Applied to Hazardous Chemical Warehouses

Storage accidents in hazardous chemical warehouses will cause great losses to people's personal safety and social economy. Therefore, real-time monitoring of the state of hazardous chemical warehouses is of great significance. Hazardous chemicals warehouse a large number of stacking similar and there is the phenomenon of occlusion, the use of image three-dimensional reconstruction can be no dead angle to monitor the warehouse state, and the point cloud registration is an important part of the three-dimensional reconstruction, a large number of similar stacking in the hazardous chemicals warehouse will cause a large number of similar structure point cloud exists, if the point cloud of the topology of the topological structure of the features is not enough to extract, it will result in the point of the mis-matching, so that the matching accuracy is reduced. Aiming at this problem, an registration method of dual-flow graph feature interaction is proposed. The registration model of dual-stream map feature interaction adopts two data streams, rotation and translation, and uses dynamic map convolution to obtain geometric features, then adds self-association feature module to extract topological features, and uses soft matching and inter-association module to obtain soft correspondences between point clouds, and carries out feature interaction based on the soft correspondences, so that the point pairs can be matched more accurately with the help of the topological structure information to solve the problem of a large number of point clouds with similar structures. The point pairs can be matched more accurately with the topology information to solve the problem of point cloud mismatching caused by a large number of similar structures. Experiments on the publicly available dataset modelnet40 and the self-constructed dataset show that this method achieves the best registration accuracy on the self-constructed dataset and the modelnet40 dataset of unseen species with fewer training rounds and achieves the second-best registration accuracy after FINet on the partial-to-partial modelnet40.