Key points are not available for this paper at this time.
In order to evacuate people safely and quickly in indoor disaster environments, it is necessary to estimate the current location of the individuals, detect disaster situations, predict disaster propagation, derive optimal individual escape paths, and implement a user-friendly and intuitive guidance system. In this study, we propose a machine-learning-based indoor augmented reality (AR) navigation and emergency evacuation system that can guide an optimal escape path for individual users. To detect emergency events and deliver sensing data, an Internet of Things (IoT) -enabled ad hoc network is considered. To deliver the sensing data safely and reliably to the server, we present a hybrid reinforcement-learning-based routing algorithm that combines direct and indirect Q -learning methods. Prediction of disaster propagation at multiple time scales is important to prevent dangerous situations. We propose a simple disaster area prediction method that ensembles elementary component gradient boosting machine models. User location is estimated by a deep neural network using the received signal strength from beacon nodes. To derive the optimum evacuation path for each individual, we propose a novel model-based Q -learning method, in which we consider the building structural model and disaster context information. The performance of the proposed system is experimentally evaluated for various disaster scenarios.
Yoo et al. (Tue,) studied this question.