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

Abstract

Largest thermal roadside dataset introduced: 11,400 annotated images and 142 video clips. Weakly supervised incremental learning framework adapts to new viewpoints without extra labelling. Achieved 8.9-point improvement in mAP on unseen data, enhancing traffic monitoring accuracy. Teacher-student network with pseudo-labelling and memory module prevents forgetting in incremental learning. Open-access dataset and code provided for reproducibility and future research. Highlights: Roadside camera systems are commonly used for traffic data collection, yet conventional optical systems are limited by poor performance in varying weather and light conditions and are often restricted by data privacy regulations. Thermal imaging overcomes these issues, enabling reliable detection across all conditions without collecting personal data. However, its widespread use is hindered by the scarcity of diverse, annotated thermal training data, especially since fixed cameras installed at the side of the road produce very similar images with the same backgrounds. This paper presents two key innovations to address these challenges: a novel dataset of 11,400 annotated images and 142 unannotated video clips, the largest and most diverse available for thermal roadside imaging to date, and a weakly supervised incremental learning framework tailored for thermal roadside imagery. The dataset supports the development of self-supervised algorithms, and the learning framework allows efficient adaptation to new camera viewpoints and diverse environmental conditions without additional labelling. Together, these contributions enable cost-effective and reliable thermal-based traffic monitoring across varied locations, achieving an 8.9-point increase in mean average precision for previously unseen viewpoints.