Abstract: The point cloud data of buildings measured by 3D laser scanning technology can clearly represent the three-dimensional spatial information of buildings, and provide high-precision and high-density surface description of buildings. The point cloud itself does not directly display its own characteristic information, so when extracting local deformation, point cloud segmentation is needed. Most of the existing segmentation algorithms applied to buildings and structures rely on the characteristics of them to set the mutation threshold. When encountering complex scenarios, these assumptions often lead to errors. With the extension of machine learning in the field of point cloud processing, a new idea has emerged in the recognition of buildings and structures' boundaries and point cloud data segmentation. Taking the 3D laser scanning data of two phases of a railway bridge above the working face of a mining area as an example, this paper uses the neural network method to segment and extract the steel structure of bridge arches. After training 10 million marked bridge points, the neural network model can learn operators to recognize the attributes of points in the point cloud and mark them, then extract the steel structure point cloud in the two phases of data. In the next way, this paper analyzes the point cloud of steel structure of bridge arches segmented by neural network, through the characteristic line fitting and length extraction of the bottom edge line of steel structure, calculates the displacement and tension of steel structure, and analyzes the bridge deformation combined with arch displacement and shape variable. The research shows that using neural network to train the labeled data can effectively identify the characteristics of buildings and structures, and this technology can be applied to analyze local deformation of them.

Key words: 3D laser scanning, point cloud segmentation, machine learning, deformation monitoring, curve fitting