[2C5-OS-7b] OS-7 (2)

[2C5-OS-7b-02] Clustering of people’s trajectories from wide-area evacuation simulation assuming a large earthquake and its application to evacuation behavior prediction

〇Takuya Oki¹, Yoshiki Ogawa² (1. Tokyo Institute of Technology, 2. The University of Tokyo)

Keywords:Evacuation behavior, Property damage, Deep learning, Clustering, Simulation

Immediately after a major earthquake, it is important to grasp the status of people staying and evacuating in real-time in order to reduce secondary damage. In this paper, it is assumed that evacuation trajectories obtained from an agent-based simulation model, which considers property damage and wide-area evacuation, as observed trajectories. First, these trajectories are classified into a small number of patterns by a sequential clustering method (DTW), and the characteristics of the obtained clusters are analyzed. Next, we discuss a method of estimating the state of evacuees based on the results of trajectory clustering. Furthermore, we construct an evacuation behavior trajectory prediction model using deep learning and LSTM, which learns the evacuation behavior characteristics of each pattern as time-series data, and evaluate the usefulness of the proposed method.