LTDv2: A Large-Scale Long-term Thermal Drift Dataset for Robust Multi-Object Detection in Surveillance

We present the Long-term Thermal Drift version 2 (LTDv2) dataset, a large-scale dataset collected using a thermal camera for object detection in a surveillance context. Collected over 9 months of continuous monitoring, LTDv2 comprises more than 1 million frames captured under diverse weather conditions, and almost 7 million annotations across four distinct classes. To provide environmental context, each video clip is enriched with detailed weather metadata, including temperature, humidity, and solar radiation. The proposed dataset is designed to support the development of algorithms and methods for outdoor recognition activities. Alongside the dataset release, we present a detailed analysis of its applications through a comprehensive case study, offering valuable insights about addressing object detection across different settings, including image enhancement and privacy-preserving anonymization. Dataset is available at 1.

Recommended citation: Parola, Marco and Aakerberg, Andreas and Johansen, Anders S and Nikolov, Ivan A and Cimino, Mario GCA and Nasrollahi, Kamal and Moeslund, Thomas B (2025). "LTDv2: A Large-Scale Long-term Thermal Drift Dataset for Robust Multi-Object Detection in Surveillance." Authorea Preprints.
Download Paper