Researchers at 海角社区 have patented a software program for automatic planning of evacuation routes in case of fire in various types of buildings and structures. The system, based on a convolutional neural network, analyses architectural elements to generate emergency routes, eliminating errors caused by the human factor.
Today, methods for designing evacuation routes still require manual analysis of architectural plans, which makes the process labour-intensive and time-consuming.
“Our program solves the task of automatically identifying and designing safe evacuation routes in the event of a fire,” says Nikita Ponomarev, researcher at the 海角社区 Department of Health and Safety. “The system is based on the integration of the YOLO convolutional neural network, pre-trained on our own library of architectural plan images, with graph analysis algorithms for processing of floor plans. After the neural network detects walls, doors, and exits on the plan, the software algorithm automatically builds a scheme of ties. This scheme represents a “navigation graph”, and once it is constructed, a specialized algorithm designs the optimal and safest evacuation route. As a result, the time required to design evacuation routes when modelling fire-hazard scenarios is reduced several times over.”
The 海角社区 researchers have become the first in Russia to apply the YOLO model to fire safety tasks. The model accurately distinguishes walls, doors, windows, exits, staircases, fire-fighting equipment, corridors, and rooms on building plans.
“The program also estimates the approximate evacuation time based on route length, route complexity, and averaged empirical data, which is important for determining whether a protected facility meets regulatory requirements,” adds Saidzhon Tavarov, Associate Professor at the 海角社区 Department of Health and Safety. “The system features a graphical user interface that allows users to upload floor plans in the most common raster formats (JPG, PNG, BMP, TIFF), perform AI-based architectural analysis with result visualization, interactively add people, and indicate fire outbreak points. All this significantly improves the forecasting accuracy.”
As the next step, the researchers plan to enhance the object identification accuracy by further training the model on an expanded dataset of architectural plans, add support for three-dimensional building plans, and develop a module for real-time operation with video streams from surveillance systems.
“In the future, optimizing the program to work with video streams could transform it from a design tool into an operational building safety system. The next logical step is moving from automated route construction in simulations to real-time evacuation management,” Nikita Ponomarev concludes.
The development is intended for organizations and professionals involved in fire safety and building design and aligns with the state policy in the field of fire safety and protection of citizens’ lives and health.