Domain Adaptation for Privacy-Preserving Pedestrian Detection in Thermal Imagery

Pedestrian detection is a core problem in computer vision, and is a problem that is gaining prominence due to its importance in assisted and autonomous driving applications. Many state-of-the-art approaches, especially those used for autonomous driving, combine thermal and visible spectrum imagery in order to robustly detect persons independent of time of day or weather conditions. In this paper we investigate two domain adaptation techniques for fine-tuning a YOLOv3 detector to perform accurate and robust pedestrian detection using thermal images. Our approaches are motivated by the fact that thermal imagery is privacy-preserving in the sense that person identification is difficult or impossible. Results on the KAIST dataset show that our approaches perform comparably to state-of-the-art approaches and outperform the state-of-the-art on nighttime pedestrian detection, even outperforming multimodal techniques that use both thermal and visible spectrum imagery at test time.