以形生行——基于图像深度学习的商业空间行为预测方法

Form Dictates Function: A Method for Predicting Commercial Space Behavior Based on Image Deep Learning

Research output: Conference PapersRGC 32 - Refereed conference paper (without host publication)peer-review

View graph of relations

Author(s)

  • 金衍孜
  • 沈彦婷
  • 谢雪莹
  • 张琪波
  • 邵帅
  • 王煦天
  • 闫超

Detail(s)

Original languageChinese (Simplified)
Number of pages10
Publication statusPublished - Nov 2024

Conference

Title2024 计算性设计学术论坛暨中国建筑学会计算性设计学术委员会年会
Location同济大学
PlaceChina
City上海
Period15 - 17 November 2024

Abstract

人类行为与建筑空间形态之间存在着本质关联。目前针对建筑空间中的人流行为预测,主要运用行人运动仿真模型,但这种方法耗时且难以反映真实环境的行为复杂性。本研究探索了一种基于图像深度学习的建筑空间人流预测方法,并以商场中庭空间为例,对其有效性进行实验验证。首先,通过智能行为感知技术,实地采集商场中庭空间的人群行为数据,并通过图像识别技术对人群数量分布进行可视化分析;进一步,将人群数量分布热力图和商场平面图进行预处理,并通过配对构建深度学习数据集;最后,通过生成对抗网络(GANs)训练模型,并通过图像生成技术实现在短时间内预测出人流热力图。该方法验证了深度学习在应对人流行为这种高度复杂现象的效果,可以在建筑设计早期阶段为空间行为推演与优化提供决策依据。
There is an intrinsic relationship between human behavior and the morphology of architectural space. Currently, pedestrian behavior prediction in architectural spaces primarily relies on movement simulation models, yet this approach is time-consuming and struggles to capture the complexity of real-world behavior. This study explores a pedestrian behavior prediction method based on image deep learning and validates its effectiveness using a case study of mall atrium spaces. First, intelligent behavior sensing technology was employed to collect on-site crowd behavior data in the mall atrium space, followed by visual analysis of the crowd distribution through image recognition technology. Next, the crowd density heatmaps and mall floor plans were preprocessed and paired to construct a deep learning dataset. Finally, the model was trained using Generative Adversarial Networks (GANs), and pedestrian flow heatmaps were predicted within a short time frame using image generation technology. This method demonstrates the effectiveness of deep learning in addressing the highly complex phenomenon of pedestrian behavior, providing decision-making support for spatial behavior simulation and optimization during the early stages of architectural design.

Research Area(s)

  • Behavioral computation, Deep learning, People flow prediction, Image recognition, Mall atriums

Bibliographic Note

Research Unit(s) information for this publication is provided by the author(s) concerned.

Citation Format(s)

以形生行——基于图像深度学习的商业空间行为预测方法. / 金衍孜; 沈彦婷; 谢雪莹 et al.
2024. Paper presented at 2024 计算性设计学术论坛暨中国建筑学会计算性设计学术委员会年会, 上海, China.

Research output: Conference PapersRGC 32 - Refereed conference paper (without host publication)peer-review