基于Grad-CAM的校园街道步行空间视觉感知体验研究  

Visual Perception Experience of Campus Street Walking Space Based on Gradient-weighted Class Activation Mapping

作　　者：李韵琴 张嘉新 谢雨辰 LI Yunqin;ZHANG Jiaxin;XIE Yuchen

机构地区：[1]南昌大学建筑与设计学院,南昌330047 [2]南昌大学城乡发展与产业创新研究院,南昌330047 [3]香港大学社会科学学院地理系,中国香港特别行政区

出　　处：《新建筑》2024年第6期18-23,共6页New Architecture

基　　金：江西省自然科学基金青年项目(20242BAB20223);南昌大学大学生创新训练计划项目(NCU-2024XC160)。

摘　　要：相较于城市街道,校园中更为简单的人群结构和低差异性的空间,提供了相对严格的控制变量条件和可重复性更高的实验环境,有助于更专注地分析特定的校园街景要素对视觉感知的影响。文章采用类激活可视化解释技术(Grad-CAM),对南方某大学的校园街道步行空间视觉感知体验进行研究。通过对比Grad-CAM类激活图与眼动追踪数据,并结合SHAP模型和核密度估计模型分析,归纳了影响校园街道空间步行感知的主要街道特征,揭示了街景要素对视觉感知的差异化影响。研究可为街道步行空间的精细化设计和更复杂的城市街道研究提供参考和借鉴。This study focuses on the pedestrian spaces of campus streets at a university in southern China,utilizing Gradient-weighted Class Activation Mapping(Grad-CAM)to explore pedestrian visual perception experiences.Grad-CAM intuitively highlights key areas in street images that affect pedestrian comfort,automatically identifying and emphasizing visual elements such as vegetation and vehicles,and revealing their impact on the pedestrian experience and visual perception.By comparing Grad-CAM activation maps with eye-tracking data,and incorporating analyses from SHAP and kernel density estimation models,the study summarizes the main street characteristics that shape pedestrian perception,and uncovers the diverse effects of streetscape elements on visual perception.This study provides architects with an analytical method that is easy to understand and apply for the refined design of campus pedestrian spaces.

关 键 词：校园街道空间 街道步行性 视觉感知体验 深度学习 类激活图 可解释性 

分 类 号：TU984.113[建筑科学—城市规划与设计]