The duality in China's traffic planning has given rise to the basic unit of urban form and function called the superblock,which is defined and bound by an arterial street network.The street network of China's ...The duality in China's traffic planning has given rise to the basic unit of urban form and function called the superblock,which is defined and bound by an arterial street network.The street network of China's superblock un derpins the coexiste nee and in teraction of global and local movement,the public and daily space,and affects place diversity and local characteristics.However,its configuration remains to be articulated because of the lack of a systematic representation method,and the associations between configuration and performance,cognition and design cannot be determi ned.This study proposes an improved representation method for the street network of China's superblocks based on Marshall's route structure analysis to explore the configurational characteristics and sustainability of the network.To fit local conditions,this study improves Marshall's route structure analysis from four perspectives,namely,the judgement of relative hierarchy,the node construction principle,and the deletion and addition of the original indicators.The improved method is then applied to calculate and compare the depth,connectivity,and complexity of the street networks of 10 sample superblocks in Nanjing,which are classified into six types by construction backgrounds,each having two seenarios differing by the level of publicity.Results indicate that the types formed in accordance with the"The Capital Plan"of the Republic of China,which presents a combination of fine orthogonal grids and radiations,and by the renewal of the traditional street-andlane network,which has the"characteristic structure" defined by Marshall,perform best in terms of configurational sustainability.The an alysis also reveals that the addition of semipublic streets formed mainly from the bottom up narrows the sustainability gap among the samples.This study provides a tool for elaborate urban study and design and provides in sights into the cognitive and practical aspects of China's urban planning and design.展开更多
点云作为一种能提供丰富空间信息与物体几何特征的数据表达形式,在自动驾驶和机器人等领域有着广泛的应用前景。点云数据具有无序性的特点,早期研究者利用深度学习工具完成点云的分类分割任务时,一般采用结构化表示方法(如将点云转换成...点云作为一种能提供丰富空间信息与物体几何特征的数据表达形式,在自动驾驶和机器人等领域有着广泛的应用前景。点云数据具有无序性的特点,早期研究者利用深度学习工具完成点云的分类分割任务时,一般采用结构化表示方法(如将点云转换成体素,多视图),但这些方法会导致部分三维空间信息损失和预处理计算量大等问题。PointNet创新性地使用共享的MLP(多层感知机)处理每个点云的向量,有效地解决了点云的无序性问题。PointNet的扩展版本PointNet++利用Encode-Decode结构进行局部特征提取。整体采用Encode-Decode结构,提出了PointRFE神经网络模型。其中,设计了能够对局部进行更充分表征的局部信息表征融合(Local Information Representing and Fusing,LIRF)模块,该模块融合了局部采样点特征,局部点云仿射变换特征和局部空间几何结构信息特征;设计了信息深度残差提取(Information Deep Residual Extractor,IDRE)模块,该模块通过带有瓶颈残差结构的共享MLP(多层感知机)对LIRF模块输出的融合特征进行深度提取。该网络模型分别在ModelNet40、ScanObjectNN和ShapeNetPart数据集上进行实验,验证网络模型在点云分类分割任务上的性能。实验结果表明,提出的网络模型在ModelNet40分类数据集整体分类精度达到93.6%;ScanObjectNN分类数据集上达到84.7%;ShapeNetPart部件分割数据集上达到86.0%。本文提出的模型在各数据集上的测试准确率均在该领域内达到先进水平。PointRFE能够充分地表征和高效地学习点云局部信息,可以很好地完成点云分类分割任务。展开更多
基金sponsored by the National Natural Science Foundation of China(NSFC#51578123)Postgraduate Research&Practice Innovation Program of Jiangsu Province(KYLX16_0234).
文摘The duality in China's traffic planning has given rise to the basic unit of urban form and function called the superblock,which is defined and bound by an arterial street network.The street network of China's superblock un derpins the coexiste nee and in teraction of global and local movement,the public and daily space,and affects place diversity and local characteristics.However,its configuration remains to be articulated because of the lack of a systematic representation method,and the associations between configuration and performance,cognition and design cannot be determi ned.This study proposes an improved representation method for the street network of China's superblocks based on Marshall's route structure analysis to explore the configurational characteristics and sustainability of the network.To fit local conditions,this study improves Marshall's route structure analysis from four perspectives,namely,the judgement of relative hierarchy,the node construction principle,and the deletion and addition of the original indicators.The improved method is then applied to calculate and compare the depth,connectivity,and complexity of the street networks of 10 sample superblocks in Nanjing,which are classified into six types by construction backgrounds,each having two seenarios differing by the level of publicity.Results indicate that the types formed in accordance with the"The Capital Plan"of the Republic of China,which presents a combination of fine orthogonal grids and radiations,and by the renewal of the traditional street-andlane network,which has the"characteristic structure" defined by Marshall,perform best in terms of configurational sustainability.The an alysis also reveals that the addition of semipublic streets formed mainly from the bottom up narrows the sustainability gap among the samples.This study provides a tool for elaborate urban study and design and provides in sights into the cognitive and practical aspects of China's urban planning and design.
文摘点云作为一种能提供丰富空间信息与物体几何特征的数据表达形式,在自动驾驶和机器人等领域有着广泛的应用前景。点云数据具有无序性的特点,早期研究者利用深度学习工具完成点云的分类分割任务时,一般采用结构化表示方法(如将点云转换成体素,多视图),但这些方法会导致部分三维空间信息损失和预处理计算量大等问题。PointNet创新性地使用共享的MLP(多层感知机)处理每个点云的向量,有效地解决了点云的无序性问题。PointNet的扩展版本PointNet++利用Encode-Decode结构进行局部特征提取。整体采用Encode-Decode结构,提出了PointRFE神经网络模型。其中,设计了能够对局部进行更充分表征的局部信息表征融合(Local Information Representing and Fusing,LIRF)模块,该模块融合了局部采样点特征,局部点云仿射变换特征和局部空间几何结构信息特征;设计了信息深度残差提取(Information Deep Residual Extractor,IDRE)模块,该模块通过带有瓶颈残差结构的共享MLP(多层感知机)对LIRF模块输出的融合特征进行深度提取。该网络模型分别在ModelNet40、ScanObjectNN和ShapeNetPart数据集上进行实验,验证网络模型在点云分类分割任务上的性能。实验结果表明,提出的网络模型在ModelNet40分类数据集整体分类精度达到93.6%;ScanObjectNN分类数据集上达到84.7%;ShapeNetPart部件分割数据集上达到86.0%。本文提出的模型在各数据集上的测试准确率均在该领域内达到先进水平。PointRFE能够充分地表征和高效地学习点云局部信息,可以很好地完成点云分类分割任务。
