A machine learning based study on pedestrian movement dynamics under emergency evacuation

Knowledge of evacuees' movement dynamics is crucial to building safety design and evacuation management. Although it is recognized that stepwise movement is the fundamental element to construct the whole evacuation process, movement pattern and its influencing factors are still not well underst...

全面介紹

Saved in:
書目詳細資料
Main Authors: Wang, Ke, Shi, Xiupeng, Goh, Algena Pei Xuan, Qian, Shunzhi
其他作者: School of Civil and Environmental Engineering
格式: Article
語言:English
出版: 2020
主題:
在線閱讀:https://hdl.handle.net/10356/143390
標簽: 添加標簽
沒有標簽, 成為第一個標記此記錄!
實物特徵
總結:Knowledge of evacuees' movement dynamics is crucial to building safety design and evacuation management. Although it is recognized that stepwise movement is the fundamental element to construct the whole evacuation process, movement pattern and its influencing factors are still not well understood. In this study, we explored the potential of adopting machine learning methods to study evacuees' stepwise movement1 dynamics based on two videos of quasi-emergency evacuation experiments. The movement patterns were categorized through Two-step Cluster Analysis and principal influencing factors were identified through Principal Component Analysis. The relationship between the movement patterns and the principal components were investigated using different modeling methods: traditional method (Multinomial Logit Model, MLM) and machine learning methods (Decision Tree, Support Vector Machine, K-Nearest Neighbor, and Artificial Neural Network). Results from two experimental videos showed reasonable consistency and the main findings are: (1) Distance to the target exit has the most pronounced effect on a single evacuee's stepwise movement pattern. (2) Surrounding evacuees' actions also have significant and complex influence on a single evacuee's stepwise movement pattern. (3) MLM showed comparable prediction accuracy with machine learning methods when the scenario is simple. The superiority of machine learning became apparent when the scenario was more complex, with a maximum enhancement of 13.25% in prediction accuracy. Each machine learning method demonstrated distinct features and advantages in different aspects.