Traffic jams are a problem in large urban areas. How can they be solved without expanding the road network? This is a challenge that researchers are attempting to address.
According to the latest global ranking by GPS manufacturer TomTom, Geneva is the city in Switzerland where commuters lose the most hours in traffic jams: 141 per year. It is among the global top 20, ahead of Zurich (116 hours) and far outstripping Lausanne (86 hours), Basel (83 hours) and Bern (43 hours). The average speed in Geneva is also Switzerland's lowest at 19.1 km/h, compared to 27 km/h in Lausanne and 42.4 km/h in Bern.
Smart traffic management is one key to improving traffic congestion, particularly in urban areas. At EPFL, the Laboratory of Urban Transportation Systems (LUTS) is using novel technology to analyze urban traffic. Drones are one such technology, and have been added to cities' existing arsenal of cameras, loop detectors and other sensors. "The problem with conventional methods is they focus mainly on cars and often have temporal and spatial limitations," says Manos Barmpounakis, a postdoc at LUTS. "But with drones, we can overcome many of the obstacles created by sensors. Drones offer a much broader, more comprehensive and more detailed view of a road network's condition."
In 2018, engineers from LUTS conducted a pioneering experiment by flying a group of drones over Athens to collect vast amounts of data and analyze traffic patterns. Because the drones couldn't distinguish license plates or faces, they complied with data protection laws. The engineers used the data to develop algorithmic methods for identifying vehicle types (cars, trucks, buses, motorcycles, bicycles, etc.) and their trajectories.
Predicting how a network will react
"We can use AI and advanced machine learning to accurately recognize, locate and track almost all pockets of congestion in wide areas," says Weijiang Xiong, a PhD student at LUTS who is studying methods for traffic congestion forecasting. His work has shown that by integrating drone measurements into classical congestion-monitoring techniques such as fixed sensors (called loop detectors), engineers can significantly improve traffic forecasting by 15% to 20% in many cases. This is crucial for designing and implementing more accurate control strategies, such as adaptive traffic signal control and signal coordination.
We won't be able to predict an accident, but if one does occur, we'll be able to predict how the network will react
"Current regulations create challenges for the technological implementation of drones, but with effective prediction methods and reliable data, we can introduce preventative measures and, for example, regulate traffic lights half an hour before a traffic jam reaches an area. We won't be able to predict an accident, but if one does occur, we'll be able to predict how the network will react," says Xiong.
To adapt models to a given city, engineers must first collect local data in order to adjust and refine the models accordingly. "While traffic congestion is a universal problem, what causes it varies from place to place. That's where the data entered into a model are crucial," says Barmpounakis. Thanks to drones, engineers can supplement existing data and deploy the technology in any city in an efficient, economical, environmentally friendly and optimal way - and thus feed more data into the model.
Measuring decibels
Drone data can also be used to analyze driving behavior while maintaining drivers' confidentiality. "We've conducted studies on lane changes and interactions between drivers," says Xiong. "Drones are the only instruments capable of providing us with this information. Individual data can show us that a vehicle braked hard, for instance, but they don't tell us why. With drones, we can see whether it's because a motorcycle cut off the vehicle, a pedestrian entered the road or the traffic light turned red."
Drones are used for safety, for analyzing multimodal traffic and driving behavior, and for measuring air quality and assessing noise pollution. This latter issue is the subject of research conducted by Jasso Espadaler Clapés at LUTS. "By knowing the kinematic profile of vehicles, such as their position, speed and acceleration, and whether they have an electric or combustion engine, we can estimate their noise and emissions," he says. "Compiling this information lets us estimate how many decibels the vehicles will produce and, with certain algorithmic models, compare this to the noise we actually hear on the street."
Taking technology from the lab to market
"Our goal is not to develop a direct solution that can be implemented overnight," says Barmpounakis. "Instead, we're studying upstream issues: to what extent can drones be useful for traffic control? What quantitative advantage does this data give us for forecasting? And what opportunities does AI provide? Practical cases will then follow."
The LUTS engineers have conducted several pilot tests, not only in Athens but also in Nairobi, Manchester and Songdo, always for research purposes. The laboratory has also given rise to a spin-off, MobiLysis, which expands the scope of implementation and real-world case studies to urban mobility systems, integrating pedestrians, active transport modes, parking and public transport into a more sustainable and human-oriented perspective.
The municipality of Pully and the canton of Geneva have already worked with MobiLysis to address their urban mobility challenges. MobiLysis has also conducted experiments in the US and is involved in major European projects. For instance, the firm measured various parameters (vehicle trajectories, speeds, acceleration, traffic flows, etc.) in Helsinki as part of the Acumen project. These data will be used to calibrate the traffic simulation software for the city's digital twin.
When mathematics helps make better predictions
While sociology can help us understand the reasons why we behave in a certain way, mathematics lets us model that behavior in order to better conceptualize, predict and prepare for future needs - including in the area of transportation. But how can we turn a quick run to the supermarket, a trip to a piano class or the daily commute into an algorithm? Conventional transportation models look only at the individual journeys people make to get from point A to point B, considering the purpose of the trip, the chosen mode of transportation and the itinerary. But engineers at EPFL's Transport and Mobility Laboratory (TRANSP-OR), headed by Michel Bierlaire, are exploring another approach - one that considers people's everyday activities (work, errands, leisure pursuits, etc.) and those of others in their household over the course of a single day as well as their entire lives.
In 2024, Janody Pougala - then a PhD student at TRANSP-OR - developed a model based on this approach. Her program accounts for people's activities and how people respond to the unpredictable events that inevitably form part of our daily lives. These factors are particularly important given today's increasingly diverse lifestyles. Commuting patterns have changed considerably, as more people work from home or carpool, and infrastructure improvements enable employees to live farther away from their employer. Pougala's model was tested successfully on a pilot prediction system at the Swiss railway company (SBB) and in an urban planning project for Zurich that involves envisioning what the city would look like if half its transportation were human-powered.
Another thing to consider is that our transportation decisions are generally made not individually, but within the context of our overall household. This is important because it means we tend to carefully plan out when and how we travel with a view to optimizing the trip. The TRANSP-OR engineers therefore incorporated household interactions into the model, enabling it to predict things like which family member gets to use the car, how household tasks are divided up, when family members are accompanied to a certain activity, when they take part in shared activities and when they turn to carpooling. The model can thus be applied to many different types of households and available transportation methods.
Our research shows that the hourly distributions produced by models calibrated at the household level better reflect actual data than those produced by models calibrated at the individual level," says Negar Rezvany, who just defended her thesis on this topic. "What's more, our model can generate realistic distributions of everyday activities and crunch through data to estimate key variables, while anchoring its calculations in behavioral theory."
Data hard to come by
One sticking point for transportation models is being able to draw on enough data - it's hard to measure people's activities due to the amount of resources required to do so, as well as for confidentiality reasons. To get around this problem, engineers use synthetic populations, which are statistically derived populations having the same characteristics as the real population. "For example, in a project with SBB, we're analyzing their data in order to build synthetic populations and use them to outline future scenarios," says Bierlaire. Other organizations interested in this approach include the Swiss Federal Office for Spatial Development, which handles issues related to Switzerland's transportation policy. Synthetic populations can also be used to model scenarios in the event of another pandemic or an economic crisis. "The trick is to find the right variables for generating predictions for spatial distributions as well as temporal ones," says Bierlaire. "Transportation habits should be evaluated over a person's entire lifetime since fundamental choices are generally made at pivotal moments during someone's life."
To that end, Rezvany analyzed urban dynamics and long-term transportation decisions in her thesis, using cross-border commuting in Luxembourg as a case study. She created a framework incorporating different time horizons: the choice of a transportation method (short term); residential relocation (medium term); and infrastructure development (long term). As Rezvany explains in her thesis: "By tracking the evolution of key indicators, the framework serves as an indicative tool to understand system behavior, anticipate future trends, and assess the long-term impacts of policy interventions."
