IIT Guwahati researchers tracked more than 36,000 vehicles using drones on an eight-lane highway near Delhi.
The study found that gaps between vehicles shrink as traffic increases but widen again near gridlock.
Six mathematical and two machine learning models were tested, with the conventional Underwood model performing best.
Researchers say India-specific traffic data could improve road design, congestion management and traffic forecasting.
Traffic on Indian roads does not always follow the patterns assumed by conventional traffic models. Two-wheelers, three-wheelers, cars, buses and heavy vehicles share road space, often moving at different speeds and changing positions according to available gaps. This diversity makes it difficult to use models developed in countries with more orderly and homogeneous traffic conditions to accurately predict how vehicles move and how congestion develops in India.
A study by researchers at the Indian Institute of Technology (IIT) Guwahati has used drone-based observations to examine these dynamics and develop traffic models that better reflect conditions on Indian roads.
The researchers recorded drone videos on an eight-lane divided highway on the outskirts of Delhi during daytime. They tracked more than 36,000 vehicles and generated 8,938,920 data points at a rate of 10 frames per second. The data were used to analyse vehicle speed, traffic density and the spacing between vehicles.
The study found that vehicles maintain larger gaps when traffic is relatively light. As traffic increases, the distance between vehicles decreases. However, when traffic becomes extremely heavy and approaches congested conditions, the gaps between vehicles begin to increase again.
The finding points to a non-linear relationship between traffic density and vehicle spacing. The distance between vehicles does not continue to decrease uniformly as more vehicles enter the road.
The researchers also reported a traffic density of around 900 vehicles per kilometre under jam conditions, measured across the full width of the road. This figure refers to the entire road section, not to an individual lane.
The study addresses a key limitation in the way traffic behaviour is modelled. Many conventional traffic models were developed using data from countries where road conditions and vehicle movements differ from those commonly seen in India. Their assumptions may not adequately capture the interactions among different vehicle types or the variations in how they move through available road space.
Speaking about the need for India-specific traffic data, Prof. Akhilesh Kumar Maurya of IIT Guwahati said, “Traffic conditions in India are different from those in many countries where existing traffic models have been developed. We need traffic data that represents the diversity and characteristics of Indian roads. Drone-based data collection provides an opportunity to observe a large number of vehicles and their movements over a road section at high spatial and temporal resolution.”
Drone-based observations allow researchers to track many vehicles simultaneously across a road section. Unlike observations taken from a single point, such data can capture vehicle movements and interactions over both space and time. This makes it possible to examine how traffic behaviour changes as road conditions become more crowded.
The researchers evaluated six mathematical traffic models and trained two machine learning models: an artificial neural network (ANN) and a random forest (RF) model.
The analysis also considered the variability in vehicle speeds under different traffic conditions. Instead of treating speed at a given traffic density as a single fixed value, the researchers developed stochastic fundamental diagrams to represent the range of speeds that may occur.
Among the conventional mathematical models tested, the Underwood model performed best and showed the most consistent results. The two machine learning models performed at a level comparable to the Greenberg model, but did not outperform the Underwood model.
The findings suggest that more complex machine learning approaches do not necessarily deliver better predictions than established mathematical models. Their performance depends on how well the models represent the traffic conditions and the data used to train them.
The researchers say that traffic models based on high-resolution, India-specific data could support more realistic traffic simulations, road and intersection design, congestion management and road-capacity assessment. They could also contribute to the development of intelligent traffic-management systems suited to Indian conditions.
Explaining the potential real-world applications, Prof. Maurya said, “Reliable traffic management begins with models that represent how traffic actually behaves on Indian roads. The high-resolution trajectory data generated through drones allows us to capture interactions among different vehicle types and the large variability that conventional traffic models often overlook. Such India-specific models can support more realistic traffic simulation, better road and intersection design, congestion management, capacity assessment and the development of intelligent traffic-management systems suited to Indian conditions.”
However, the findings should be interpreted in the context of the study's scope. The analysis was based on daytime observations from a single eight-lane divided highway near Delhi. Traffic behaviour can vary significantly across road types, cities, weather conditions and time periods. Further observations from a wider range of Indian roads will be needed to establish how well these models perform in different settings.
The research paper, titled Regression-based Macroscopic and Microscopic Analysis of UAV-Based Data of a Disorderly Stream, has been published in the International Journal of Intelligent Transportation Systems Research.
