| 摘要: |
| 城市地标可视性研究是城市设计及景
观风貌保护等相关规划的重点内容。目前，传统
的“眺望”视线控制方法在可视域划定、可视度
分析上存在缺陷，后续结合数字化手段改进的
三维可视域模拟分析方法部分解决了可视域划
定的问题，但数据精度要求高，且无法满足可视
度分析要求。针对此问题，本文提出了一种基于
迁移学习与智能模式识别的城市地标可视性分
析方法，以南京紫峰大厦为例，利用自主采集的
街景数据，结合地标数据集，训练出结合迁移
学习、深度神经网络的人工智能体，完成对不同
尺寸下，符合紫峰大厦特征的地标识别。通过改
进的智能模式识别方法，可以实现地标的可视域及可视度识别。经验证，分析结果较过去的“眺望”视线控制方法，三维可视域模拟分析方法
更为精准、真实。弥补现有方法在可视域、可视度分析上的不足。 |
| 关键词: 地标可视性 人工智能 迁移学习 模式识别 |
| DOI：10.13791/j.cnki.hsfwest.20240302 |
| 分类号: |
| 基金项目: |
|
| Urban landmark visuality research based on transfer learning and intelligent patternrecognition: A case study of Nanjing Zifeng Tower |
|
XU Yunyi,ZHANG Shengyue,JIANG Jinliang,CHEN Wenlong
|
| Abstract: |
| Urban landmark refers to the iconic structures or landscapes in the city, which are the
imagery elements in the city with the significance of direction guidance, form unification, value
symbolization, historical remembrance and other spatial significance. Cities have paid great attention to
the sightline protection and control of landmarks in the hope of highlighting the city’s iconic image.
Existing research on landmark building sightline protection and control focuses on two
directions, one is the method of landmark building sightline protection planning, such as landscape
view corridor control, sightline zoning control, fuselage overlook landscape control, etc., basically
adopting the “overlook” method of sightline control, i.e., a number of viewpoints are selected by
the planner or the local government in the city, and it is required that the viewpoints are located at a
certain point of view. The planner or local government selects a number of viewpoints in the city, and
requires that the landmark buildings can be viewed from the viewpoints without being negatively
affected by the neighboring buildings or environmental elements on the overall scene of the
landmark buildings. However, the early landmark line of sight analysis is in plan view, delineating
the relatively regular sectors and spindle shapes for control, which is more idealized and lacks the
refined line of sight simulation analysis. Therefore, with the development of geographic information
technology, there is a second major research direction, which is the visual domain simulation analysis
of technical methods, such as GIS visual domain analysis, WebGL three-dimensional visual domain
analysis, Cesium three-dimensional visual domain analysis, etc., these methods are from the threedimensional
space to start, through the geographic information data simulation analysis, improve the
fineness of the line of sight analysis. However, the method’s data quality requirements can’t always
be fulfilled, according to the current commonly used data accuracy is difficult to meet the demand
for visuality analysis, such as building height often can not be obtained directly, often with a floor
height of 3 meters projected, such as terrain elevation accuracy of 10 meters as a unit loaded into the
system for the remittance of the analysis. In addition, the data of greening, utility poles and other
structures are difficult to obtain, as these mentioned above are not loaded into the system such as
GIS as an influence factor, resulting in the accuracy of the visuality analysis is limited. Therefore,
there is an urgent need for a technology that reflects the visuality of landmark buildings from a real
perspective and assists in landmark sightline protection planning and control.
Artificial intelligence-based streetscape recognition technology expands the methods of urban
environment analysis and provides a new opportunity for landmark visuality analysis. With the
advantages of wide coverage, high data refinement, low data collection cost, and intuitive and real
scenes, streetscape images have become an important data source for urban environment evaluationresearch. On the other hand, the continuous breakthroughs in artificial intelligence technology have greatly improved the efficiency of identification and
evaluation of urban built environment based on streetscape images on a wide range of spatial scales, and the authenticity and fineness of environmental
evaluation have also been improved. However, in general, there is a lack of research on the visuality of specific landmark buildings in streetscape using
artificial intelligence, and the main difficulty lies in the accurate identification of specific landmark buildings of different sizes in streetscape images, and the
construction of methods and training of models still need to be further explored.
On the basis of existing research, this paper proposes a landmark building visuality analysis method based on migration learning and intelligent pattern
recognition. Taking the Nanjing landmark building Zifeng Tower as an example, more than 6200 street scene images of Nanjing Gulou District are acquired
by panoramic camera, and the GLDv2 (Google Landmark Datasets Version 2.0, GLDv2) is utilized to calibrate the size and construct the training set to
realize the recognition of different sizes of the Zifeng Tower in each image, and the model learning process is carried out through migration learning method
and deep neural network (DNN). The model is loaded into the artificial intelligence body to perform pattern recognition of different sizes of Zifeng Tower
appearing in the images, and obtain the visuality of Zifeng Tower on all roads in Gulou District. Compared with traditional geographic information methods
such as GIS, street view is closer to the real perspective of individuals, and can provide a real scene of “seeing is believing”. Comparative study found that the
artificial intelligence analysis method of streetscape images, based on the GIS method, further identifies the influence factors such as greening and structure
shading, micro-topography, etc., and significantly improves the identification accuracy, correcting 40% of the GIS analysis results. Applying the AI analysis
results to urban planning and design can identify streets with potential for enhancement and suggest street enhancement for landmark landscape effects. |
| Key words: landmark visuality artificial intelligence transfer learning pattern recognition |
