An Unsupervised Transfer Learning Framework for Visible-Thermal Pedestrian Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Dual cameras with visible-thermal multispectral pairs enable 24/7 pedestrian detection for autonomous driving and intelligent transportation systems.
• Detector performance degrades significantly in new environments due to domain shift, necessitating large, manually annotated datasets.
• Manual data annotation is time-consuming, labor-intensive, and not scalable for real-world applications.

Purpose of the Study:

• To develop a novel unsupervised transfer learning framework for multispectral pedestrian detection.
• To adapt existing multispectral pedestrian detectors to new target domains without requiring manual annotations.
• To improve the scalability and efficiency of pedestrian detection systems in diverse real-world scenarios.

Main Methods:

• Proposed an unsupervised transfer learning framework utilizing pseudo training labels for domain adaptation.
• Employed auxiliary detectors and adaptive label fusion strategies based on estimated environmental illumination.
• Generated intermediate domain images via source-to-target image translation to enhance model pre-training.

Main Results:

• Achieved new state-of-the-art performance on the KAIST and FLIR ADAS datasets.
• Demonstrated effective domain adaptation without any manual training annotations on the target data.
• Validated the framework's ability to generalize across different datasets and environmental conditions.

Conclusions:

• The proposed unsupervised transfer learning framework significantly enhances multispectral pedestrian detection performance.
• Eliminating the need for manual annotation in target domains makes the approach highly scalable and cost-effective.
• This method offers a robust solution for deploying reliable pedestrian detection systems in various autonomous applications.