基于点云目标检测算法的船体分段合拢面构件识别方法

Component Recognition Method of Block Erection Surface Based on Point Cloud Object Detection Algorithm

在船体分段合拢面精度检测方面,三维激光扫描仪相比全站仪具有高效、高精度和操作便捷等优势。但扫描点云往往数据量庞大且会包含大量与设计模型无关的周围空间点,不仅增加运算时间而且影响配准精度。基于显著构件的点云配准方法能解决该问题,但实现显著构件的智能识别,还需要一种针对船体分段合拢面构件的智能识别算法。采用深度学习方法,构建一种基于点的、无锚点单阶段目标检测神经网络模型,其适用于船体分段合拢面点云数据,基本实现了对船体分段合拢面上构件的智能识别。使用ADAM优化器对网络进行优化训练,在测试集上获得了平均精确度均值P_(A-m)为64.36%的效果。研究成果可用于改进点云粗配准方法,为实现船体分段合拢面精度的智能高效检测提供帮助。

In terms of the accuracy detection of block erection surface,3D laser scanners have advantages over total stations in terms of efficiency,accuracy,and ease of operation.However,scan-generated point clouds often entail vast amounts of data,frequently including numerous surrounding spatial points unrelated to the design model.This not only extends computational time but also undermines registration accuracy.The point cloud registration method based on salient components effectively addresses this problem.In order to achieve intelligent recognition of salient components,a smart recognition algorithm specific to block erection surface components is required.A deep learning method is adopted to build a point-based,anchor-free,single-stage object detection neural network model that is suitable for block erection surface point cloud data and basically achieves intelligent recognition of components on the block erection surface.The network is optimized and trained using the ADAM optimizer and a P_(A-m) of 64.36%is achieved on the test dataset.The results can be used to improve the coarse registration method of point clouds and provide assistance in achieving intelligent and efficient accuracy detection of block erection surfaces.