An Artificial Intelligence and Computer Vision Framework for Construction Safety Evacuation Modeling Across Complex Building Layouts

When a building is finished, its safety plan is usually a static document: a drawing showing where the doors are and a calculation of how long it would take everyone to leave if an alarm sounded. This approach works well for permanent structures where the layout never changes and the number of people inside is predictable. But construction sites are different. They are living, shifting environments where walls go up and down, stairwells remain unfinished, and the number of workers changes with every shift. In these chaotic spaces, the old rules often fail because they cannot account for a blocked hallway or a sudden crowd of workers moving through a narrow corridor. To keep people safe, engineers need a way to test evacuation plans not just on paper, but in a dynamic model that understands how real people move through changing spaces. This requires combining the ability to read building drawings with the power to simulate how crowds react when things go wrong.

A researcher at Arizona State University, Bhargav Vaghani, has developed a new system called PeopleFlow to solve this exact problem. Instead of relying on fixed assumptions, this framework uses computer vision to automatically read floor plans and turn them into a digital playground where virtual workers can move around. The system does not try to predict exactly where every single person will go, which is impossible, but it does simulate how groups of people behave when they are in a hurry. By running thousands of these simulations, the research reveals that the time it takes to clear a building can vary wildly depending on the shape of the rooms and the strategy people use to find their way out. In some layouts, the difference in evacuation time between a good plan and a bad one is more than eight times.

The core of this work is a pipeline that turns a simple image of a floor plan into a complex simulation. First, the system looks at a drawing of a building and automatically identifies the walls, doors, and exits. It then builds a map that the computer can use to navigate. Inside this digital map, the system creates hundreds of virtual agents, each representing a worker. These agents move according to a set of rules that mimic real human behavior: they try to reach an exit, they avoid bumping into each other, and they react to crowding. If a path becomes too crowded, the agents can choose to switch to a different door, just as a real person would if they saw a long line forming. The researchers tested this system on nine different types of building layouts, ranging from long school corridors to large shopping malls, and ran over 650 separate experiments. They changed the conditions in each test, sometimes blocking an exit, sometimes adding more people, and sometimes making the virtual workers act as if they were panicking.

The results showed that the shape of a building matters far more than many safety codes assume. In one test, a layout with long, straight corridors took nearly 183 seconds to clear under normal conditions, while a large, open hall with many exits cleared in just 24 seconds. This massive difference highlights that a building's geometry is the most critical factor in safety. The study also found that how people choose their route makes a huge difference. When the virtual workers were programmed to simply head for the closest door, they often created massive bottlenecks, leading to long delays. However, when they were allowed to look at the crowd and choose a less busy exit instead, the time to clear the building dropped by nearly 80 percent in some cases. This suggests that in a real emergency, clear signage or instructions that encourage people to spread out could save significant time.

The research also explored what happens when things go wrong. When a primary exit was blocked in the simulation, the time to evacuate some buildings increased dramatically, with one layout taking up to 350 seconds to clear. This confirms that having a single point of failure in a building's design can be catastrophic. Interestingly, the study found that making people move faster does not always help. When the virtual workers were made to move at higher speeds, they arrived at the exits faster than the doors could handle, causing more congestion and actually slowing down the overall evacuation. This counterintuitive finding suggests that in a panic, the urge to run might make the situation worse by creating a jam at the door.

What makes PeopleFlow particularly useful for construction safety is its ability to adapt to the changing nature of a job site. Because the system can automatically read new floor plans, a safety manager could upload a drawing of the building as it looks today, run a simulation, and see if the current layout is safe. If the simulation shows a bottleneck forming at a specific corner, the manager could test a solution immediately, such as adding a temporary exit or moving a pile of materials, and see if the problem disappears. The system also connects to the idea of a "digital twin," a virtual copy of a real building that updates in real time. In the future, cameras on a construction site could feed live data about where workers are located directly into the simulation. This would allow the system to warn managers if a specific area is becoming too crowded and to suggest the best evacuation route based on the actual conditions at that moment.

The researchers are careful to note that this is a simulation, not a crystal ball. The model does not account for smoke, fire, or the complex psychological effects of true panic, and it currently works only in two dimensions, meaning it does not yet fully simulate moving up and down stairs. However, the framework provides a powerful tool for comparing different designs and strategies. It moves safety planning away from guessing and toward evidence-based decisions. By showing that small changes in layout or routing can have massive effects on safety, the work offers a way to design buildings that are safer before the concrete is even poured. The ultimate goal is to give site managers a way to test their plans against the worst-case scenarios, ensuring that when an alarm sounds, the path to safety is clear, efficient, and ready for the people who need it.