Overcoming Data Scarcity in Roadside Thermal Imagery: A New Dataset and Weakly Supervised Incremental Learning...

Area of Science:

• Computer Vision
• Machine Learning
• Transportation Engineering

Background:

• Conventional roadside cameras struggle with variable weather, lighting, and privacy concerns.
• Thermal imaging offers a solution for reliable, privacy-preserving traffic data collection.
• Limited diverse, annotated thermal data hinders widespread adoption of thermal imaging for traffic analysis.

Purpose of the Study:

• To address the scarcity of thermal roadside imaging data.
• To develop a robust learning framework for thermal imagery analysis.
• To enable cost-effective and reliable thermal-based traffic monitoring.

Main Methods:

• Creation of the largest and most diverse thermal roadside imaging dataset to date (11,400 annotated images, 142 video clips).
• Development of a weakly supervised incremental learning framework tailored for thermal roadside imagery.
• Utilizing the dataset to support self-supervised algorithms and framework adaptation to new viewpoints and conditions.

Main Results:

• The novel dataset and framework facilitate efficient adaptation to new camera viewpoints and environmental conditions.
• Achieved an 8.9-point increase in mean average precision for previously unseen viewpoints.
• Demonstrated the potential for cost-effective and reliable thermal-based traffic monitoring.

Conclusions:

• The developed dataset and learning framework significantly advance thermal imaging applications in traffic monitoring.
• This approach overcomes limitations of traditional optical systems and data scarcity challenges.
• Enables enhanced traffic analysis across diverse scenarios without compromising privacy.