无监督迁移学习红外行为识别

Unsupervised Transfer Learning Infrared Human Action Recognition

为了缩小可见光视频和红外视频之间的模态差异而提高红外行为识别率,以及简化深度学习用于红外行为识别需人工标注数据集繁琐的问题.基于迁移学习的思想,本文提出一种用可见光动作(源域)来识别红外动作(目标域)的无监督异构红外行为识别算法(UHDIAR).UHDIAR算法将可见光数据和红外数据映射到同一个对齐的特征空间中,采取余弦相似度调整源域样本的权重,利用对齐后的可见光数据训练权重支持向量机(W-SVM),进而识别红外动作并自动标注.采用可见光动作数据集(XD145)和红外动作数据集(InfAR)进行实验,结果表明UHDIAR的平均识别率与标准的SVM相比相对提高68.65%.

In order to improve the recognition rate of infrared human action recognition by narrowing the modal difference between visible light video and infrared video,and simplify the tedious problem of manually annotating dataset when infrared human action recognition is recognized by deep learning.Motivated by the idea of transfer learning,we propose an unsupervised heterogeneous domain infrared action recognition(UHDIAR)algorithm that uses visible light action(source domain)to identify infrared motion(target domain).UHDIAR algorithm maps the visible light data and the infrared data into the same aligned feature space,adjusts the weight of source domain samples with cosine similarity,and uses the aligned visible light data to train the weight support vector machine(W-SVM),so as to identify infrared action and automatically label.Visible light action dataset(XD145)and infrared action dataset(InfAR)were used for experiments.The experimental results show that the average recognition rate of UHDIAR is 68.65%higher than that of the standard SVM.