2024 – Present
This project aims to tackle the challenge of ineffective decision-making during active shooter incidents in public places, such as schools. Life-threatening mistakes occur when crucial evacuation decisions are made with incomplete or flawed information during high-stress situations. The primary objective is to develop an end-to-end, real-time public guidance and advisory tool to improve crisis decision-making capabilities for indoor environments.
The project utilizes Graph Neural Networks and AI-based Deep Reinforcement Learning models to model indoor environments as Graph Networks. These models are used to compute optimal evacuation strategies in real-time. Data sources include graphical representations of real-life school layouts and Unreal Engine-based simulations to validate the AI model’s performance.
The project seeks to integrate its findings into real-world scenarios by providing a decision support system that complements existing security measures. This system will aid key decision-makers, such as school personnel and first responders, in planning and executing optimal evacuation strategies, thus enhancing safety and improving emergency response effectiveness in a crisis. The project incorporates two under-explored approaches to its research: utilizing advanced AI techniques to compute real-time evacuation routing under a dynamic threat (such as a gunman) and integrating architectural aspects like safe rooms and door actuation into the evacuation strategy.
In 2024, the project achieved several milestones, including developing a capacity-constrained optimization technique and a deep reinforcement learning model that made the optimization suitable for real-time implementation. A risk-aware dynamic lockdown algorithm was formulated to trap shooters without creating excessive risk imbalances across the building; this work was presented at the 2024 SRA Annual Meeting in Austin, TX. Additionally, the team developed a digital twin using Unreal Engine 5 for full verification and validation of shooter detection and egress algorithms. Key collaborations were also established with other SENTRY research thrusts to enhance data collection and model refinement, incorporating crowd behavior models and evaluating the effects of potential architectural modifications.
This project works closely with SENTRY’s Real-Time Video Surveillance for Threat Detection and Mitigation project and the School Security Case Study. These collaborations help with data collection and simulation testing, validating models with real-world data and testing environments and ensuring that the developed solutions are relevant and practical for deployment.