Method of Point Cloud Segmentation at Component Level of Railway Bridges Based on Improved RandLA-Net
With the rapid development of UAV and laser scanning technology,point cloud data has become an important information source for the construction progress management of railway bridges.Refined segmentation of bridge structure at the component level is a critical step in monitoring and assessing construction progress.Therefore,a method of point cloud segmentation at the component level of railway bridges based on improved RandLA-Net was proposed.First,the self-attention mechanism was considered to capture the global context semantic relations between discrete points and embed position encoding blocks to enhance the position features of points.Then,weights or pooling operations were dynamically adjusted according to input features to improve the attention pooling module so that the network can focus on specific areas of input data and adaptively aggregate information from adjacent points.According to the results,this method yielded a mean Intersection over Union(mIoU)2.4％higher than that of RandLA-Net in processing the point cloud dataset of railway bridges,and it delivers satisfactory segmentation performance.
3D point clouddeep learningsemantic segmentationrailway engineering dataattention mechanismpoint cloud data
NETL
NSTL
万方数据
