关键词: 步态识别, 卷积神经网络, 深度学习, GaitSet, 损失函数

Abstract: Gait recognition, as an emerging biometric recognition technology, has broad application prospects in fields such as crime prevention, forensic identification, and social security. Although sequence based methods can preserve more spatiotemporal information, they are computationally expensive and lack flexibility. To overcome the limitations of these methods, this paper proposes an optimized GaitSet gait recognition algorithm with a refinement module to optimize the structure of the original network. Adding normalization operations after the convolutional layer can accelerate the convergence speed of the network. Introducing SA attention mechanism improves the performance and generalization ability of the model. By adopting joint loss training, this paper uses the Softmax loss function to compensate for the instability, slow convergence, and easy overfitting in the model training process caused by the triplet loss function. By using CASIA-B data for experiments, the paper converts the gait data into an energy map, which is more conducive to extracting more information during the training process of the network. The experiments show that the optimized GaitSets algorithm has high recognition accuracy for all angles on the CASIA-B dataset.

Key words: gait recognition, convolutional neural networks, deep learning, GaitSet, loss function