低分辨率红外感知系统的人体行为数据增强

针对目前公开的人体行为红外数据集较少的问题,本文设计了基于生成对抗网络和基于3D人体红外模型的两种数据增强方法,以快速扩充现有的红外人体行为数据集。搭建了改进的生成网络模型AC-WGAN并添加了网络优化策略,以生成高质量的红外热图;利用Unity 3D引擎搭建表征红外特征及运动信息的人体模型,模拟红外阵列传感器成像原理,实现了给定人体模型及传感器方位信息便能自动生成大量且多样化数据的功能。基于数据增强后的数据集训练和测试卷积神经网络,实验结果表明,网络对不同行为的感知准确度达到80%以上,且显著提高了网络对陌生数据的识别能力,所设计的数据增强方法对于扩充人体行为红外数据集有效。

Human behavior data augmentation for the low-resolution infrared perception system

To solve the problem of the lack of public human behavior datasets, two data augmentation methods are designed, based ongenerative adversarial networks and 3D human infrared models, to rapidly expand existing infrared human behavior datasets in thisarticle. An improved generative network model AC-WGAN is formulated by adding network optimization strategy to generate high-qualityinfrared heat maps. The Unity 3D engine is used to build a mannequin with infrared features and motion information. By simulating theimaging principle of infrared array sensors, the function is realized to automatically generate a large and diverse amount of data for agiven mannequin and sensor orientation information. A convolutional neural network is established, which is based on the data-enhanceddataset. Experimental results show that the perceptual accuracy of different behaviors is up to 80% . The ability of the network to identifyunfamiliar data is significantly improved and the effectiveness of the designed data augmentation method is proven for expanding thehuman behavior infrared dataset.