Human detection with smoke occlusion based on AI-generated images

Abstract

Detecting occupants trapped in buildings is challenging during fire accidents because of the dense, rapidly spreading smoke. To this end, this paper proposes a human detection method that accounts for smoke occlusion, based on artificial intelligence (AI)-generated images and an improved real-time detection transformer (RT-DETR) model. First, a procedure is designed to construct a dataset of human images with smoke occlusion using generative AI, addressing the current dataset shortage. Second, an RT-DETR model is optimized using a transformer-based global-local interaction network to increase the sensitivity to human characteristics with smoke occlusion. Third, corresponding validations were conducted. The results indicated that the proposed method could achieve remarkable average precision (up to 93.8 %) for smoke occlusion scenarios. Finally, the proposed method was employed in real-world fire cases. The proposed method can accurately detect humans with smoke occlusion in building fires, effectively aiding fire emergency evacuation and rescue efforts.