融合轮廓增强和注意力机制的改进 GaitSet步态识别方法

针对传统基于轮廓的步态识别方法受限于输入特征及模型特征提取的能力，从而导致识别准确率不高的问题，提出一种融合轮廓增强和注意力机制的改进GaitSet步态识别方法。首先通过预处理获取行人的轮廓图，求得其均值，合成步态GEI能量图，将其作为神经网络模型的输入特征，增强了人体外观的表示。其次在提取特征的过程中引入注意力机制，增强模型的特征提取能力，从而提高步态识别的精度。最后在CASIA-B和OU-MVLP数据集上进行实验，所提方法的平均Rank-1准确率分别为87.7%和88.1%。特别是在最复杂的穿大衣行走条件下，相较于GaitSetv2算法，准确率提升了6.7%,表明所提出方法具有更强的准确性。此外，所提方法几乎没有增加额外的参数量、计算复杂度和推理时间，说明其各模块的快速性。

Improved GaitSet method for gait recognition via fusion of silhouette enhancement and attention mechanism

Aiming at the problem that traditional gait recognition methods based on silhouette are limited by the ability to extract input features and model features, which leads to low recognition accuracy, an improved GaitSet method for gait recognition via fusion of silhouette enhancement and attention mechanism is proposed. Firstly, the outline of the pedestrian is obtained by preprocessing, and its average value is obtained. Then the GEI energy diagram is synthesized, which is used as the input feature of the neural network model to enhance the representation of human appearance. Secondly, the attention mechanism is introduced in the process of feature extraction to enhance the feature extraction ability of the model, so as to improve the accuracy of gait recognition. Finally, experiments are carried out on the CASIA-B and OU-MVLP benchmark data sets, and the average Rank-1 accuracy of the proposed method is 87.7% and 88.1%, respectively. Especially under the most complex walking conditions with overcoat, compared with GaitSetv2 algorithm, the accuracy is improved by 6.7%, indicating that the proposed method has stronger accuracy. Notably, the proposed innovative method adds almost no additional parameter number, computational complexity, and inference time, which proves the rapidity of its individual modules.