In view of the limitations of traditional measurement methods in the field of building information,such as complex operation,low timeliness and poor accuracy,a new way of combining three-dimensional scanning technolog...In view of the limitations of traditional measurement methods in the field of building information,such as complex operation,low timeliness and poor accuracy,a new way of combining three-dimensional scanning technology and BIM(Building Information Modeling)model was discussed.Focused on the efficient acquisition of building geometric information using the fast-developing 3D point cloud technology,an improved deep learning-based 3D point cloud recognition method was proposed.The method optimised the network structure based on RandLA-Net to adapt to the large-scale point cloud processing requirements,while the semantic and instance features of the point cloud were integrated to significantly improve the recognition accuracy and provide a precise basis for BIM model remodeling.In addition,a visual BIM model generation system was developed,which systematically transformed the point cloud recognition results into BIM component parameters,automatically constructed BIM models,and promoted the open sharing and secondary development of models.The research results not only effectively promote the automation process of converting 3D point cloud data to refined BIM models,but also provide important technical support for promoting building informatisation and accelerating the construction of smart cities,showing a wide range of application potential and practical value.展开更多
针对当前两阶段的点云目标检测算法PointRCNN:3D object proposal generation and detection from point cloud在点云降采样阶段时间开销大以及低效性的问题,本研究基于PointRCNN网络提出RandLA-RCNN(random sampling and an effectivel...针对当前两阶段的点云目标检测算法PointRCNN:3D object proposal generation and detection from point cloud在点云降采样阶段时间开销大以及低效性的问题,本研究基于PointRCNN网络提出RandLA-RCNN(random sampling and an effectivelocal feature aggregator with region-based convolu-tional neural networks)架构。首先,利用随机采样方法在处理庞大点云数据时的高效性,对大场景点云数据进行下采样;然后,通过对输入点云的每个近邻点的空间位置编码,有效提高从每个点的邻域提取局部特征的能力,并利用基于注意力机制的池化规则聚合局部特征向量,获取全局特征;最后使用由多个局部空间编码单元和注意力池化单元叠加形成的扩展残差模块,来进一步增强每个点的全局特征,避免关键点信息丢失。实验结果表明,该检测算法在保留PointRCNN网络对3D目标的检测优势的同时,相比PointRCNN检测速度提升近两倍,达到16 f/s的推理速度。展开更多
文摘In view of the limitations of traditional measurement methods in the field of building information,such as complex operation,low timeliness and poor accuracy,a new way of combining three-dimensional scanning technology and BIM(Building Information Modeling)model was discussed.Focused on the efficient acquisition of building geometric information using the fast-developing 3D point cloud technology,an improved deep learning-based 3D point cloud recognition method was proposed.The method optimised the network structure based on RandLA-Net to adapt to the large-scale point cloud processing requirements,while the semantic and instance features of the point cloud were integrated to significantly improve the recognition accuracy and provide a precise basis for BIM model remodeling.In addition,a visual BIM model generation system was developed,which systematically transformed the point cloud recognition results into BIM component parameters,automatically constructed BIM models,and promoted the open sharing and secondary development of models.The research results not only effectively promote the automation process of converting 3D point cloud data to refined BIM models,but also provide important technical support for promoting building informatisation and accelerating the construction of smart cities,showing a wide range of application potential and practical value.
文摘针对当前两阶段的点云目标检测算法PointRCNN:3D object proposal generation and detection from point cloud在点云降采样阶段时间开销大以及低效性的问题,本研究基于PointRCNN网络提出RandLA-RCNN(random sampling and an effectivelocal feature aggregator with region-based convolu-tional neural networks)架构。首先,利用随机采样方法在处理庞大点云数据时的高效性,对大场景点云数据进行下采样;然后,通过对输入点云的每个近邻点的空间位置编码,有效提高从每个点的邻域提取局部特征的能力,并利用基于注意力机制的池化规则聚合局部特征向量,获取全局特征;最后使用由多个局部空间编码单元和注意力池化单元叠加形成的扩展残差模块,来进一步增强每个点的全局特征,避免关键点信息丢失。实验结果表明,该检测算法在保留PointRCNN网络对3D目标的检测优势的同时,相比PointRCNN检测速度提升近两倍,达到16 f/s的推理速度。
