视觉基础模型支持的老城区更新规划方法与实践  

Old District Renewal Planning Method and Practice Based on Vision Foundation Model

作　　者：梁程程 卓文淖 左琛 张军飞 LIANG Chengcheng;ZHUO Wennao;ZUO Chen;ZHANG Junfei

机构地区：[1]陕西省城乡规划设计研究院 [2]西安市城市规划设计研究院 [3]长安大学运输工程学院 [4]道路基础设施数字化教育部工程研究中心 [5]西安建筑科技大学建筑学院

出　　处：《规划师》2024年第4期88-97,共10页Planners

基　　金：国家重点研发计划项目(231111520200)。

摘　　要：对于具有复杂建成环境的老城区,采用人工智能技术可以提高更新规划的科学性,然而现有的目标识别方法高度依赖人工标注数据集,面临更新实践不足、分析精度受限和训练成本过高的局限。将视觉基础模型引入城市更新领域,探讨视觉基础模型支持的老城区更新规划方法:通过等间隔采样构建图像块数据集,使用视觉基础模型提取高维视觉特征;启动余弦距离、流形学习、高斯混合模型等机器学习技术,实现图像块数据集的智能分组;对图像块进行空间转译,精准识别重点区域;结合多维要素综合评判,优化智能识别结果,指导后续更新规划实践。同时,以陕西省榆林市清涧县老城区为例,探索该技术方法在城市更新规划实践中的应用,以期为老城区的城市更新提供新的思路。Artificial intelligent technology may improve the planning rationality in complex built environment of old districts.However,current object detection methods are highly dependent on annotated datasets,with limitations including insufficient renewal practices,limited analysis accuracy,and high training cost.Old district renewal planning method is discussed by introducing vision foundation model:an image patch dataset is built by evenly spaced sampling,and high-dimensional visual features are extracted with the vision foundation model;then a range of machine learning techniques is activated including cosine distance,manifold learning,and Gaussian mixture model to realize image pstch organization;next,the image patches are spatially translated to recognize key areas;finally,combined with comprehensive determination,the image detection result is improved to guide the subsequent renewal planning practice.The proposed method is tested in a case study in Qingjian county,Yulin city,Shaanxi province with promised efficiency and accuracy,providing a new thinking for old district renewal.

关 键 词：视觉基础模型 城市更新 老城区 卫星遥感图像 陕西省榆林市清涧县 

分 类 号：TU984[建筑科学—城市规划与设计]