Transformer fusion-based scale-aware attention network for multispectral victim detection

• Published in: Complex & intelligent systems Vol. 10; no. 5; pp. 6619 - 6632
• Main Authors: Chen, Yunfan, Li, Yuting, Zheng, Wenqi, Wan, Xiangkui
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.10.2024; Springer Nature B.V; Springer
• Subjects: Attention; Complexity; Computational Intelligence; Computer science; Convolutional neural network; Data Structures and Information Theory; Datasets; Deep learning; Disasters; Drone aircraft; Drones; Engineering; Evacuations & rescues; Intelligent systems; Low visibility; Multispectral images; Natural disasters; Neural networks; Original Article; Pedestrians; Smart drones; Smog; Thermal imaging; Transformers; Unmanned aerial vehicles; Victim detection; Deep learning; Victim detection; Convolutional neural network; Smart drones; Multispectral images
• DOI: 10.1007/s40747-024-01515-y
• ISSN: 2199-4536

The aftermath of a natural disaster leaves victims trapped in rubble which is challenging to detect by smart drones due to the victims in low visibility under the adverse disaster environments and victims in various sizes. To overcome the above challenges, a transformer fusion-based scale-aware attention network (TFSANet) is proposed to overcome adverse environmental impacts in disaster areas by robustly integrating the latent interactions between RGB and thermal images and to address the problem of various-sized victim detection. Firstly, a transformer fusion model is developed to incorporate a two-stream backbone network to effectively fuse the complementary characteristics between RGB and thermal images. This aims to solve the problem that the victims cannot be seen clearly due to the adverse disaster area, such as smog and heavy rain. In addition, a scale-aware attention mechanism is designed to be embedded into the head network to adaptively adjust the size of receptive fields aiming to capture victims with different scales. Extensive experiments on two challenging datasets indicate that our TFSANet achieves superior results. The proposed method achieves 86.56% average precision (AP) on the National Institute of Informatics—Chiba University (NII-CU) multispectral aerial person detection dataset, outperforming the state-of-the-art approach by 4.38%. On the drone-captured RGBT person detection (RGBTDronePerson) dataset, the proposed method significantly improves the AP of the state-of-the-art approach by 4.33%.