Can AI Fix Road Congestion? IISc's UVH-26 Dataset Brings Indian Reality To Traffic AI

Bengaluru: Traffic congestion remains one of the most persistent urban challenges in India, draining productivity, increasing fuel consumption, and worsening air pollution. While AI is increasingly being deployed under Smart City and urban mobility initiatives, many of these systems struggle to deliver results on Indian roads. The reason lies in the data they are trained on.

Most traffic AI models used in India rely on Western datasets that assume disciplined lane behaviour and limited vehicle diversity. Indian roads, in contrast, are dominated by two-wheelers, auto rickshaws, buses, and a wide variety of informal and mixed-use vehicles. This mismatch causes AI systems to misclassify vehicles or miss them entirely, undermining applications such as signal optimisation, congestion detection, urban planning, and safety analysis.

UVH-26: A Dataset Designed for Indian Traffic

To address this gap, the AI for Integrated Mobility (AIM) initiative at the Indian Institute of Science (IISc) launched UVH-26 in November 2025, which is India’s largest open-source traffic image dataset. Released alongside a set of AI models, UVH-26 is built entirely from Indian roads, containing 26,646 high-resolution images collected from 2,800 Safe City CCTV cameras across Bengaluru, with 1.8 million annotated bounding boxes covering 14 Indian Road Congress–defined vehicle categories.

The dataset was created through a national Urban Vision Hackathon that brought together 565 student volunteers. To ensure accuracy, annotations were generated using consensus-based techniques, producing high-quality labels suitable for real-world deployment. Unlike global datasets such as COCO, UVH-26 captures the density, diversity, and unpredictability of Indian urban traffic.

Why Detection Accuracy Is the Foundation of Traffic AI

Professor Yogesh Simmhan, Associate Professor at IISc’s Department of Computational and Data Sciences, in an exclusive interview with ETV Bharat, pointed out that accurate detection and classification are the foundational building blocks of all urban mobility applications. Vehicle counting, congestion analysis, adaptive signal control, and road safety monitoring all depend on correctly identifying what is actually moving on the road.

Example base images (left) and their pre-annotations with bounding box and label (right) (Special Arrangement)

When these foundational tasks fail, even the most advanced AI systems become ineffective. While modern AI architectures are already powerful, their real-world success depends largely on the data used to train them. By providing representative Indian traffic data, UVH-26 enables AI systems that are far more accurate and reliable in Indian conditions.

India-Specific Models Show Significant Accuracy Gains

Along with the dataset, AIM@IISc has released six fine-tuned vehicle detection models based on YOLOv11, DAMO-YOLO, and RT-DETRv2 architectures. These models show performance improvements of up to 31.5 per cent compared to standard COCO-trained baselines when evaluated on Indian traffic.

Professor Simmhan cautioned that these gains vary by vehicle type. For globally common vehicles such as sedans, improvements are modest. However, for India-specific vehicles—including auto rickshaws, buses, trucks, and varied SUV types—the accuracy gains are substantial. These improvements directly address the categories that dominate Indian roads and are most critical for traffic management.

The following table summarises the technical highlights of UVH-26:

How Better AI Translates Into Decongested Roads

In Indian cities, congestion is rarely caused by a lack of road space alone; it is more often the result of inefficient signal timing and poor visibility into real traffic conditions. This is where accurate, India-trained AI begins to make a measurable difference.

Teams training models during the on-campus Urban Vision Hackathon (Special Arrangement)

Precise vehicle classification allows traffic systems to estimate actual traffic volumes and lane saturation more reliably, especially in mixed traffic conditions where two-wheelers and auto rickshaws are frequently undercounted. When integrated into AI-driven traffic systems such as adaptive signal control and congestion prediction, these improvements in accuracy deliver long-term, cumulative benefits.

Even seemingly modest gains—such as improving detection accuracy from 70 per cent to 85 per cent—may save only a minute or two per trip for an individual commuter. However, when multiplied across millions of daily journeys, these small time savings add up to reduced fuel consumption, lower emissions, improved productivity, and smoother, more predictable urban mobility.

Bengaluru as a Real-World Testbed

By using AI models trained on local traffic conditions, Bengaluru’s traffic managers are now better equipped to study congestion patterns over time and test targeted interventions at high-pressure junctions. Bengaluru has emerged as a pilot city for this initiative, backed by over three years of collaboration between AIM@IISc and the Bengaluru City and Traffic Police.

The police enabled access to Safe City camera imagery, allowing researchers to test AI models under real operating conditions. This collaboration demonstrates how academic research can directly support city-level traffic management and planning. The insights gained in Bengaluru can be replicated across other Indian cities facing similar congestion challenges.

Performance comparison of UVH-26 trained models against state-of-the-art baselines (Special Arrangement)

Opportunities for Smaller Cities

Professor Simmhan noted that smaller cities can still benefit from UVH-26 because Safe City cameras are far more common than traffic cameras. Even in Bengaluru, traffic cameras are limited and mainly cover major roads, while Safe City cameras—with wider coverage and for safety, but not to track traffic—exist across most cities and towns. Although safety-camera data is harder to analyse, even a small number of such cameras can be useful. By running these models continuously, cities can collect long-term data to understand traffic patterns and evaluate changes, such as one-way roads, through before-and-after analysis.

The key, Professor Simmhan emphasised, is a scientific, methodical approach—working closely with transportation engineers and using AI as a decision-support tool.

UVH-26 to improve signal timing and reduce congestion or accidents

Simmhan explained that cities usually depend on vendors for AI traffic solutions, and institutions cannot deploy systems citywide on their own. However, cities can use these models to raise standards:

• First, in tenders, cities should require vendors to match or exceed the performance of the free, open models released by AIM@IISc. This immediately raises the benchmark.
• Second, cities should use India- or city-specific benchmarking datasets to objectively evaluate vendor models. Vendors must demonstrate accuracy on local data, enabling fair, scientific comparison. This will push vendors to either improve their models or adopt the open ones, improving solution quality across the board.
• Finally, cities should measure real impact. Running time-bound 3-6 months pilots and measuring real-world outcomes—such as congestion levels and junction throughput—can ensure that AI deployments deliver tangible results rather than remaining experimental.

Since its release on 5 November, UVH-26 crossed 1,500 downloads, becoming the top-trending open-source object detection dataset on Hugging Face for nearly two weeks. This response highlights strong demand among Indian students, startups, and researchers for India-specific training data. Until now, most traffic AI models in India were trained and validated on foreign datasets, limiting their effectiveness on Indian roads. UVH-26 marks a shift toward AI that is grounded in Indian reality.

Balancing Innovation With Privacy

Because UVH-26 is derived from CCTV footage, privacy and ethical considerations were central to its development. All sensitive metadata, including camera identifiers, locations, and timestamps, was removed, and visible faces and license plates were masked. At the same time, the team avoided excessive blurring that could reduce model accuracy, carefully balancing privacy protection with usability.

UVH-26, India’s largest open traffic dataset aims to power AI for Indian roads (Getty Images)

What Lies Ahead

Looking ahead, AIM@IISc plans to release multi-city datasets through biannual data challenges, covering diverse conditions such as low-light environments, peak traffic, and seasonal variations. Sponsorship support from industry is valuable in enabling these crowdsourced efforts, tapping into student enthusiasm. Parallel work is underway on traffic safety analytics, including identifying accident-prone blackspots and benchmarking road safety compliance. The team is also exploring “what-if” scenario analysis, such as understanding traffic patterns during large events like IPL matches or assessing the impact of ride-hailing vehicle parking. Bengaluru will continue to serve as the pilot city, with successful solutions scaled to other Indian cities.

Open Data for National Impact

By releasing UVH-26 under a Creative Commons license and the trained models under permissive open-source licenses, AIM@IISc has made both public and commercial use possible at no cost. This open-access approach enables cities and vendors across India to deploy better traffic solutions immediately.

UVH-26 lays the foundation for AI-driven mobility systems that can genuinely reduce congestion, improve road safety, and make Indian cities more livable—using data built in India, for India. For the everyday commuter, the change may not be dramatic overnight. But fewer unnecessary stops, smoother intersections, and more predictable travel times are precisely the incremental gains that, when scaled across millions of journeys every day, can quietly transform how Indian cities move.