应急场景下无人机多光谱数据融合的人体识别方法

Human body identification method for multispectral data fusion of unmanned aerial vehicles in emergency scenarios

针对传统测绘无人机在应急安防等领域中人体识别研究数据短缺,识别效率低下,存在误检、漏检的问题,该文模拟复杂地理环境进行无人机多光谱数据采集,基于Unet的编解码结构提出了一种融合多光谱特征的轻量级双分支网络(PMS-Unet)。利用并行的轻量化卷积神经网络(MobileNetv3)骨干网络以及空间通道挤压激励(scSE)注意力机制,同时结合遥感指数构建特征输入和改进损失函数来提高人体目标的识别率。该文的模型在自组织数据集进行人体目标提取实验,并与PSPNet、Unet++、DeepLabV3+和RTFNet进行对比。结果表明,该文模型能够有效提升人体识别效果,在应急模拟场景下鲁棒性较好。

Aiming at the shortage of personnel identification research data,the false detection and missed detection of the identification efficiency of traditional surveying and mapping UAVs in the emergency security and other fields,the complex geographical environment for UAV multispectral data collection was simulated and a lightweight dual-branch network PMS-Unet that integrated multispectral characteristics was proposed based on Unet’s codec structure in this paper.A parallel MobileNetv3 backbone network and the scSE attention mechanism combined with the remote sensing index were used to construct feature inputs and improve the loss function and the recognition rate of human targets.The proposed model was compared with PSPNet,Unet++,DeepLabV3+and RTFNet by human target extraction experiments in the self-organizing dataset.The results showed that the study integrated multispectral features and could effectively improve the human body recognition effect,which was robust in emergency simulation scenarios.