Vision-Based Air Quality Estimation for Smart Cities

TRAQID dataset and AQIFormer for scalable, sensor-free AQI monitoring

Overview

Urban air quality monitoring traditionally depends on dense sensor deployments, which are expensive, sparse, and difficult to maintain.

This project explored camera-based AQI estimation as a scalable alternative, integrating traffic imagery and environmental context to support smart city governance.

The work was conducted at IIIT Hyderabad (SPCRC Lab) under the broader initiative on IoT-enabled smart cities.

Problem

  • AQI sensors are costly and require maintenance.
  • Sparse deployment limits spatial resolution.
  • Pollution hotspots are difficult to monitor in real-time.
  • Cross-city generalization remains challenging.

We asked:

Can visual traffic data combined with environmental signals estimate AQI reliably?

TRAQID – Traffic-Related Air Quality Image Dataset

To enable research in vision-based pollution modeling, we created:

TRAQID

  • 26,678+ traffic-related images
  • Co-located AQI and meteorological data
  • Multi-condition coverage (day, night, varying traffic density)
  • Designed for real-world deployment scenarios

Presented as a Spotlight Paper at ICVGIP 2024 and among the most downloaded papers of the conference.

AQIFormer – Transformer-Based Multi-View Architecture

We proposed AQIFormer, a transformer-based model that integrates:

  • Multi-view traffic imagery
  • Meteorological inputs
  • Context-aware feature fusion

Key Results

  • 90% AQI classification accuracy
  • 88% performance during night conditions
  • 81.67% cross-city transfer accuracy
  • Strong generalization across domains

These results demonstrated that camera-based AQI estimation can serve as a low-cost, scalable alternative to sensor-heavy deployments.

Contributions

  • Designed end-to-end dataset pipeline (image + AQI + weather integration)
  • Proposed multi-view transformer-based architecture
  • Conducted cross-city generalization experiments
  • Demonstrated robustness across lighting conditions

Impact

  • Enables low-cost environmental intelligence
  • Supports smart city planning and governance
  • Reduces dependency on dense sensor infrastructure
  • Bridges Computer Vision and IoT systems
  • TRAQID – ICVGIP 2024 (Spotlight)
  • AQIFormer – ICVGIP 2024
  • Protocol for Hunting PM2.5 Emission Hot Spots – EnvSys 2023

References