基于生成对抗网络的多视角步态识别研究

Multi-View Gait Recognition Using Generative Adversarial Network

在步态识别过程当中,易受到拍摄角度、行人着装、携带物遮挡等因素的影响,从而降低了识别精度。为了缓解这些协变量因素的影响,多位学者提出将任意步态图转化到固定视角的方法,但在需要识别的步态图之间的角度差异过大时,转化过程会导致步态特征的丢失。为了更好的保留步态图的特征信息,提出了一种多视角的识别模型。模型使用生成对抗网络的方法,将任意状态的步态图转化为多角度正常状态下的步态图集合,从而尽可能多地保留原本的特征信息,来提升识别精度。最后在CASIA-B数据集上进行了测试和验证,结果表明在选取合适角度组合时,识别率相较之有较明显的提高。

Gait recognition is easily affected by the view of angle,clothing,presence of bag and so on during the process of recognition,which can adversely affect the recognition accuracy.In order to mitigate the influence of these covariates,previous studies proposed a method to transform arbitrary gait view to a fixed view.However,when the difference of views is too large,the transformation process will loss more gait characteristics.This paper proposes a multi-angle recognition model to address this problem.The model can be used to generate a set of gait images in the normal condition with multiple angles for an arbitrary gait image,so that the gait feature information can be retained as much as possible to achieve a better recognition effect.Experiments on CASIA-B dataset show that the recognition rate of proposed method is significantly improved when the appropriate angle combination is selected compared with the previous research,indicating that the model does realize the retention of more feature information and has certain feasibility.