基于双流步态网络的跨视角步态识别

通过增强样本数据和网络特征,提出双流步态网络,增强模型对携带物、衣物变化影响的鲁棒性.首先构造双流步态网络,分别提取步态视频数据中的全局特征和协变量影响范围外的局部判别信息.再将两组网络的特征信息相加融合后,得到步态的双流特征表达.提出的限制随机遮挡策略增广用于训练样本的难度和多样性,提高网络对局部特征的学习能力,减弱协变量的不利影响.另外,改进三元组损失采样方法,加速网络模型的训练收敛速度.在大型步态数据集CASIA-B和OU-MVLP上的实验表明,在携带背包和穿着不同衣物的行走状态下,双流步态网络步态识别准确率较高.

Two-Stream Gait Network for Cross-View Gait Recognition

With data augmentation and network feature map augmentation, a two-stream gait network is proposed to enhance the robustness of the model against the influence of belongings and clothing variations. Firstly,both global features and local discriminative information in gait videos are extracted by two-stream network. Then, the representation of gait feature is obtained by integrating outputs of two streams. The proposed restricted random mask is utilized to promote the network to learn more discriminative features and reduce the influence of belongings and clothing variations simultaneously. Furthermore, a triplet loss sampling algorithm is improved to accelerate the training convergence speed of the network model. Experiments on datasets, namely CASIA-B and OU-MVLP, indicate that the proposed method achieves a high gait recognition accuracy under different bagging and clothing walking conditions.