[2H11-14] Physical Protection

[2H14] Temporal scene graph based action sequence analysis for malicious behaviors identification in nuclear security

*ZHAN LI¹, XINGYU SONG¹, SHI CHEN¹, KAZUYUKI DEMACHI¹ (1. UTokyo)

Keywords:nuclear security, action sequence analysis, deep learning, temporal scene graphs

Malicious behaviors identification in nuclear facilities is a complex task, which requires accurate analysis of human action sequences. The prevailing malicious behaviors identification approaches are incomprehensive in the process of action sequence analysis, which consider only pre-defined malicious actions while ignoring the relationship between actions and other factors such as objects and background. In this paper, a novel framework is proposed to perform comprehensive human action sequences analysis based on temporal and human-object actions. First, we employ deep learning-based action recognition model and object detection model to obtain temporal actions and human-object actions, respectively. Subsequently, the human action sequences are structured as temporal scene graphs for malicious identification and visualized for secondary verification.