Abstract: For slice 3D point cloud,the building of consecutive slices was interpreted as animated curves tracing,and an unsupervised neural network was proposed,which incrementally reconstructed curves from scattered points.The solution firstly re-sampled point cloud by slicing plane to yield 2D samples,and then presented them to the network in succession for growing,shrinking,moving and reconnecting dynamically,which extracted local principal components and topologies of samples.Upon convergence on one slice,the network represented several polylines as current slice output.Meanwhile,those polylines served as initial estimation for next slice.The re-sampling-learning routine iterated overall subsequent slices to construct them incrementally.Compared with existing non-incremental approaches,the proposed solution could accelerate point cloud slicing taking advantage of shape coherence between finely spaced slices and reconstruct hundreds or thousands of topologically correct slices without user intervention.

Key words: additive manufacturing, curve reconstruction, point cloud slicing, unsupervised neural network

摘要： 针对三维点云分层问题,将连续层片构造过程解释为对运动曲线的追踪,提出一种非监督神经网络用于从散乱点中增量式重构各层曲线。首先用分层平面对点云重采样得到二维样本点,然后逐个输入网络,使其动态增长、删除、移动、改变连接,从而抽取样本的局部主成分和拓扑结构。网络在当前层学习收敛后,其结构表达了一组多义线,该组多义线同时作为下一层的初始估计,迭代执行重采样—学习过程,实现了连续层片的增量构造。算例表明,与现有非增量分层方式相比,所提算法可有效利用层间相似性提高分层效率,并在无用户干预的前提下构造成百上千层拓扑正确的层片。

关键词: 增材制造, 曲线重建, 点云分层, 非监督神经网络