Detection of Suspicious Behaviors based on Intention Inferred from Human Trajectories Using Inverse Reinforcement Learning

Abstract

Detecting suspicious behaviors from behavioral records is important for preventing crime and trouble. However, detection by observers requires huge work and experiential skill, so automatic detection methods are required. Existing methods regard behaviors with low appearance probability as suspicious. However, in these settings, behaviors that are not actually suspicious but have low appearance probability are erroneously regarded as suspicious. Considering this, the purpose of this paper is automatic suspicious behaviors detection that does not require suspicious behaviors data, is capable of real-time detection, and is more accurate than existing methods. We regard behaviors that does not meet the "normal" intention as suspicious, and use Inverse Reinforcement Learning to infer intention from human trajectories. Experiments using self-made trajectories confirmed that our proposed method could accurately evaluate behaviors that existing methods incorrectly did. In addition, we evaluated with evaluation metrics, and confirmed that our proposed method showed high performance.