CN 11-5366/S     ISSN 1673-1530
“风景园林,不只是一本期刊。”

基于定向视觉追踪的公共空间多主体行为计算分析方法

Computational Analysis Method for Multi-subject Behavior in Public Spaces Based on Targeted Computer Vision Tracking

  • 摘要:
    目的 针对儿童人群的行为量化分析是城市公共空间研究的新需求。构建基于定向视觉追踪的多主体行为计算分析方法,可以弥补常规方法难以甄别不同人群类型的问题,揭示儿童、家长与空间的协同交互规律,支撑面向具体人群行为规律的空间优化设计。
    方法 以公共空间多主体行为交互规律为研究对象,采用“技术研究—方法构建—实例论证”的研究路径,首先构建定向目标行人追踪的技术框架,然后探究针对不同人群交互规律的计算分析与可视化方法,最后以儿童游憩公共空间为例,验证多主体行为计算分析方法。
    结果 通过对中心放射型和线性带状两类儿童游憩空间的比较分析,揭示出多主体行为计算方法的3个关键应用效果:基于人体比例特征的追踪技术可以在公共空间尺度实现目标儿童人群的识别;基于聚集程度、静态使用率、动态使用率的计算分析可以系统解析成人与儿童的交互关系;基于可视化热力图的交叉比较分析可以揭示空间特征对多主体交互行为的干预原理。
    结论 多主体行为计算分析方法可以为多种人群交互的空间行为研究提供支撑,可应用于复杂人群场景的公共空间使用后评估和优化设计。

     

    Abstract:
    Objective Quantitative analysis of children’s behaviors has emerged as a new requirement in the research on urban public spaces. To this end, the primary focus of research lies in the precise spatiotemporal positioning of the crowd within the space. Traditional methods often involve long-term video recording and manual notation of crowd positions in every frame using the “observation method”. While effective, these methods are time-intensive. With the development of computer vision technology, it has become possible to automate the tracking of complex crowd behaviors in public spaces, thereby introducing novel methodologies for computational analysis in urban public space. However, the detailed identification and analysis of different crowd categories, such as children and parents, remain a significant challenge. This research aims to establish a computational analysis method for multi-subject behavior based on targeted computer vision tracking. This method reveals interaction patterns among children, parents and spatial morphology, thereby supporting spatial optimization designs for specific crowd behaviors.
    Methods Taking multi-subject interaction behaviors in public spaces as the research object, this research adopts a three-stage research approach: technological investigation, methodology construction, and case study validation. Initially, the technical framework for targeted pedestrian tracking is established. Video data is recorded from selected angles based on spatial conditions, ensuring adequate representation of spatial-temporal dynamics. And a pre-trained deep learning model is adopted for precise localization and trajectory annotation of pedestrians. Subsequently, computational analysis and visualization methods for revealing the interaction behaviors of different groups are explored, which involves a pedestrian identification threshold model based on human proportion characteristics that enables targeted identification and differentiation of children from adults, and a dedicated analysis module designed to visualize behavioral patterns of each identified crowd and thereby provide visual patterns for the spatial-temporal distributions of different crowds. Finally, the effectiveness of the multi-subject behavior analysis framework is validated through a case study on children’s recreational public spaces. The research selects two typical children’s recreational public spaces located in commercial areas. It analyzes three key behavioral metrics: average spatial distance distribution, stay duration distribution, and passer-by count distribution. Correlation analyses and interpretations of these metrics reveal the interaction patterns between children and parents and their relationship with the spatial morphological layout.
    Results The computational analysis method for multi-subject behavior enables long-term, large-scale behavioral data collection and analysis for different crowd categories. The case study on the children’s recreational public spaces reveals that, children’s activities in radially organized spatial layouts tend to be concentrated independently in central areas, while parents often move along the periphery for supervision. No significant overlap between the activity areas of children and parents is observed, suggesting minimal need for spatial overlap consideration. In such designs, the focus should be on the orientation of children’s activity spaces, as the layout of play facilities affects the observation points of supervising parents. In linear spatial layouts, parents and children closely accompany each other, primarily engaging in stationary supervision. These layouts require the consideration of spatial overlap between parents and children, as well as additional seating or rest facilities. The placement of play facilities in linear spaces significantly influences both children’s resting positions and parents’ supervision points. The empirical findings indicate that tracking technology based on human proportion features is effective for identifying target children and adults crowds at the scale of public space. The computational analysis method based on congestion degree, static usage rate, and dynamic usage rate systematically reveals adult − child interaction dynamics, and the cross-comparative analysis using visualized heatmaps uncovers the effects of spatial features on multi-subject interaction behaviors.
    Conclusion The computational analysis method for multi-subject behavior supports spatial behavioral research involving interactions of various crowd categories and is applicable to post-occupancy evaluations and design optimization in complex public spaces. It facilitates targeted spatial renovations and facility placements based on the actual spatial usage and behavioral requirements of different crowds. The research further recognizes existing technological limitations and potential future developments. While the method effectively differentiates adults and children using body aspect ratios, it cannot yet distinguish other demographic groups and their detailed semantic behaviors. Therefore, future development using human pose tracking is essential for more refined analysis. Furthermore, this research primarily explores technical methodologies based on a case study on children’s recreational spaces in commercial areas, resulting in certain sample limitations. Future research should expand the case categories, propose comprehensive optimization principles, and validate outcomes through feedback from practical projects.

     

/

返回文章
返回