On Indian roads, where cars, buses, three-wheelers, and two-wheelers are often forced to share limited space, traffic can change drastically in mere seconds. For urban planning specialists and traffic engineers, understanding this dynamic is a complex task because many traditional models assume that vehicles strictly follow lanes and maintain predictable speeds, which rarely corresponds to reality on Indian roads.
Researchers from the Indian Institute of Technology (IIT) Guwahati are working to bridge this gap. They have developed mathematical models that better account for the unpredictable behavior of transport on Indian roads by combining drone video footage, automatic vehicle tracking, and machine learning.
Tracking the positions of nearly nine million vehicles
The research team, led by Professor Akhilesh Kumar Maurya from the Department of Civil Engineering, studied traffic movement on an eight-lane divided highway on the outskirts of Delhi. Using drones equipped with high-resolution cameras, they recorded about five hours of daytime traffic in 4K resolution. These videos were then processed using DataFromSky software, which extracts vehicle movements from aerial imagery.
The result was a dataset comprising 36,818 vehicles and 8,938,920 data points on vehicle positions, captured at a frequency of 10 frames per second. The researchers categorized the vehicles into six groups: two-wheelers, three-wheelers, cars, buses, heavy commercial vehicles, and light commercial vehicles. This allowed for a detailed picture of how different types of transport moved, how closely they followed each other, and how their speed changed as traffic density increased.
Professor Maurya emphasized the need to use India-specific data, explaining that 'road conditions in India differ from the conditions in many countries where existing traffic models were developed. We need traffic data that reflects the diversity and characteristics of Indian roads.'
Drone-based data collection provides the ability to observe a large number of vehicles and their movements on a road section with high spatial and temporal detail.
The researchers found that the distance between vehicles varied depending on traffic density. On relatively free roads, vehicles maintained larger intervals, but as traffic grew, these intervals decreased. However, under severe congestion, the arrangement patterns changed again.
This variability is important because two vehicles moving under seemingly similar conditions will not necessarily move at the same speed. To understand these differences, the team tested six traffic mathematical models, including established approaches such as Greenshields and Underwood. Additionally, they trained two machine learning models: Artificial Neural Networks (ANN) and Random Forest (RF).
The most interesting aspect was the creation of stochastic fundamental diagrams. Simply put, traditional traffic models usually estimate one speed for a given traffic density. Stochastic models go further by considering the range of speeds at which vehicles can actually move under those conditions. This allows the models to represent the randomness of Indian traffic more realistically, instead of smoothing out the differences between individual vehicles.
From improving forecasts to smarter road planning
The researchers found that the Underwood model showed the greatest stability among the tested mathematical models, achieving an R² value of 0.86. Both machine learning models showed approximately 0.83.
The results obtained can find practical application in assessing highway capacity, designing intersections, determining road speeds, and improving traffic modeling. The data can also help researchers calibrate car-following models that study how drivers regulate their distance and speed relative to the vehicles ahead.
Professor Maurya noted: 'Reliable traffic management starts with models that reflect how traffic actually behaves on Indian roads. High-detail trajectory data obtained via drones allows us to capture the interaction between different types of vehicles and the greater variability that traditional traffic models often overlook.'
Such India-specific models can support more realistic traffic simulation, improve road and intersection design, manage congestion, assess capacity, and develop intelligent traffic control systems adapted to Indian conditions.
For anyone who has spent time on the roads in India, the complexity of traffic forecasting is easily understood: a car slows down, a two-wheeler enters a narrow gap, an auto-rickshaw changes lanes, and the movement of hundreds of vehicles can change in seconds. By meticulously recording these interactions, the IIT Guwahati team hopes to make traffic models more reflective of what passengers experience daily. This research is a step toward understanding not only how many vehicles use the road, but also how they move relative to each other, which can ultimately help make the daily commute more predictable.