基于多通道步态集合的跨视角步态识别

为了提高步态识别的识别效率,提出了一种基于多通道步态集合的步态识别算法.算法中将一个周期的步态轮廓图进行归一化处理后,通过步幅的大小将周期内的轮廓图划分为7个通道,每个通道内的轮廓图按照平均步态来处理,得到7通道的步态模板.结合卷积神经网络的优异性能,将该多通道步态模板(period energy im-age,PEI)当成图像集作为网络的输入,让网络自身去提取步态模板之间的时序关系.在此基础上,使用增强约束的Triplet Loss函数来学习到最优的特征和距离度量,使最终的训练更加有效.通过在标准数据集CASIA-B上的实验结果表明,与之前多数算法相比,所提出的算法在跨视角的情况下能取得很好的识别效果.

Cross-view gait recognition based on multi-channel gait set

In order to improve the efficiency of gait recognition, this paper proposes a gait recognition algorithm based on multi-channel gait set. After normalizing the silhouettes of a cycle in the algorithm, the silhouette of a period are divided into seven channels by the size of the stride, and the silhouette of each channel is processed according to the average gait. Finally, the period energy image of seven channels is obtained. This paper combines the excellent performance of convolutional neural network, regards the multi-channel gait template period energy image (PEI) as the input of the network, and lets the network itself extract the temporal relationship between the period energy images. On the basis of these, the Triplet Loss function with enhanced constraints is used to learn the optimal features and distance metrics, making the final training more efficient. The experimental results on the standard data set CASIA-B show that compared with most previous algorithms, the proposed algorithm can achieve good recognition effect in the case of cross-view.