Simulation-Based Machine Learning for Detection of Rarely Occurring Behaviors Based on Human Trajectories

Abstract This study focuses on the identification of rarely occurring human behaviors that could pose a potential security threat in public spaces, especially relevant to autonomous robotic bodyguards. The goal is to develop machine learning models capable of recognizing behaviors indicative of hostile intent. Rarely occurring behaviors in this context refer to unusual movement patterns, such as loitering with suspicious intent, sudden directional changes toward a protected individual, or evasive maneuvers following detection. Since such events are rare in real-world data, we propose a simulation-based approach to generate synthetic trajectories mimicking these behaviors, which are then used to train robust recognition models. Our findings demonstrate that a combination of simulated and real-world data improves the detection of these behaviors, making autonomous security systems more effective in mitigating potential threats. Examples of test results from our models can be found online here.