A Strong Baseline for Multi-Person Tracking in Thermal Infrared Imagery

Multi-person tracking is a crucial component in many computer vision solutions for autonomous driving or surveillance related tasks. While extensive research exists in the visible spectrum, the applicability of established multiperson tracking approaches to thermal infrared images is largely unexplored, despite its high relevance for practical applications. This work investigates the importance of commonly used tracking modules for detection, motion modeling, person re-identification, and association in the thermal domain. On the basis of our findings, we develop a strong multi-person tracker for thermal imagery, which significantly outperforms the baseline method of the novel Thermal MOT dataset (+15.6MOTA, +22.5 IDF1). With comprehensive experiments, differences to tracking on data in the visual spectrum are revealed, and our single tracking components are explored in detail. Moreover, tackling the limitations of many existing methods for real-time applications, we develop a runtime-optimized version of our tracker, which runs at 81 FPS, while still achieving state-of-the-art results. This work presents our approach to the PBVS Thermal Pedestrian Multiple Object Tracking Challenge 2025. The code is publicly available on GitHub: https://github.com/StadlerDaniel/ThermalTrack.