基于深度学习的步态识别方法

Gait recognition method based on deep learning

步态识别是一种非侵入性生物识别技术,可用于在监控系统中识别身份。然而协变量因素如视角、服装、携带物等使步态识别性能下降,其中视角是一项具有挑战性的因素。卷积神经网络(CNN)是最先进的机器学习技术之一,具有拟合复杂非线性函数的能力。因此,提出一种基于多层CNN的步态识别方法。该方法使用CNN直接从低级输入原始数据(即步态能量图(GEI)),自主学习步态特征,利用四重损失函数对网络进行端到端的训练。应用上述方法在具有挑战性的CASIA-B数据集上进行步态识别测试。实验结果显示,该方法在单视角和交叉视角的条件下识别精度分别达到93.78%和91.68%,对降低步态识别性能的几种因素具有鲁棒性,在实际应用中具有一定潜力。

Gait recognition is a non-invasive biometric technology that can be used to identify pedestrians in surveillance systems.However,the performance of gait recognition is challenged due to covariate factors,such as detection perspective,clothing,carrying items and so on.Convolution Neural Network(CNN)is one of the most advanced machine learning techniques and has the ability to fit complex nonlinear functions.Therefore,a cross-view gait recognition method based on multi-layer CNN was proposed.The method used CNN to directly input raw data(ie,Gait Energy Image(GEI))from a low level,learned gait characteristics autonomously,and used the quadruple loss function to train the network end-to-end.The above method was applied to cross-view gait recognition on challenging CASIA-B data sets.The experimental results show that the proposed method is robust to several factors that reduce gait recognition performance and has some potentialities in practical application.