显著性时空特征融合的多视角步态识别算法

Fusion of Salient Spatio-Temporal Features Based Multi-View Gait Recognition

行人步态是一种具有唯一标识行人身份能力的生物特征,可以使用视频监控系统中的行人步态远距离识别行人身份,如何有效提取监控视频中行人的多视角步态特征是一个具有挑战性的问题。本文设计了一个孪生LSTM网络架构,用于完成多拍摄视角下的步态识别任务。本文方法优点:①设计了一个框架融合步态序列显著的空间特征和时序特征进行多视角步态识别;②设计了时序总结孪生LSTM架构自动学习不同视角下步态序列的显著周期性运动特征;③分析了LSTM的层数、隐藏单元数量与识别准确率的关系,定量评价了相同视角和交叉视角下的步态识别性能。实验结果表明,本文方法在OULP-C1V1-A步态数据集上相同视图和交叉视角下都取得了良好的步态识别性能。

Gait is a biometric feature with the ability to uniquely identify a person,which can be used to identify a person from others at a distance in video surveillance systems.How to effectively extract the multi-view gait features is a challenge problem.In this paper,a siamese LSTM network architecture is designed for multi-view gait recognition.Compared with existing gait recognition methods,our proposed approach has several advantages:①A framework is designed to fuse the salient spatio-temporal features of gait sequences for multi-view gait recognition;②A temporal summary siamese LSTM architecture is proposed to automatically learn the salient periodic features of multi-view gait sequences;③The relationship between the number of LSTM layers,the number of LSTM hidden units and the recognition accuracy is analyzed,and the performance of the intra-view and cross-view gait recognition is quantitatively evaluated.Experimental results show that the proposed method achieves much better performance in intra-view and cross-view matching on the OULP-C1V1-A gait dataset.