Scholarlink

<p>This article studies the methods and models of gait recognition in recent years and makes certain improvements to improve accuracy and practicality by addressing their shortcomings. Even though the models for gait recognition are being upgraded every year, the accuracy is still unsatisfactory, especially for outdoor datasets. Even worse, the vast majority of models rely heavily on the batch size during training, which will affect the performance and feedback time. The main reason is that a single graphics cared has too limited computing power to process a large amount of data simultaneously. This manuscript studies the methods and models of gait recognition and makes certain improvements based on the state-of-the-art model. The filter response normalization (FRN) layer achieves normalization without relying on batch-dimension statistics; therefore, it can reduce the model’s sensitivity to the batch size, and we replaced the batch normalization layer in the residual convolutional neural network with the FRN layer innovatively. Due to the absence of an average removal component in the standardization of FRN, a TLU function has been leveraged with a threshold mechanism as the activation function, which can enhance the nonlinear expression capability of the model. At the same time, a more complex and reasonable global pooling function has been used to extract spatial features of gait. This function optimizes the global feature extraction strategy and enables more effective capture of spatial information in gait data. Finally, we utilize the high-performance computing mode in the CUDA environment for training and processing data, and the real-time performance has been improved significantly. The new model has been named: new GaitBase. In the study, we conducted experiments on multiple common gait recognition datasets, such as CASIA-B, GREW, and Gait3D. The results reflect that New GaitBase achieved significant results beyond previous gait recognition models (2023CVPR: GaitBase) under all conditions. The method achieves a significant improvement in Rank-1 accuracy compared to existing benchmarks: it outperforms CASIA-B by 2.4%, GREW by 29.1%, and Gait3D by 26.4%. Code is available at: <a href="https://github.com/yigemeinianda/New_GaitBase_version1.0">https://github.com/yigemeinianda/New_GaitBase_version1.0</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

New GaitBase: better practicality in gait recognition without batch dependence

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Haoran Yu,
  • Li Wang

摘要

This article studies the methods and models of gait recognition in recent years and makes certain improvements to improve accuracy and practicality by addressing their shortcomings. Even though the models for gait recognition are being upgraded every year, the accuracy is still unsatisfactory, especially for outdoor datasets. Even worse, the vast majority of models rely heavily on the batch size during training, which will affect the performance and feedback time. The main reason is that a single graphics cared has too limited computing power to process a large amount of data simultaneously. This manuscript studies the methods and models of gait recognition and makes certain improvements based on the state-of-the-art model. The filter response normalization (FRN) layer achieves normalization without relying on batch-dimension statistics; therefore, it can reduce the model’s sensitivity to the batch size, and we replaced the batch normalization layer in the residual convolutional neural network with the FRN layer innovatively. Due to the absence of an average removal component in the standardization of FRN, a TLU function has been leveraged with a threshold mechanism as the activation function, which can enhance the nonlinear expression capability of the model. At the same time, a more complex and reasonable global pooling function has been used to extract spatial features of gait. This function optimizes the global feature extraction strategy and enables more effective capture of spatial information in gait data. Finally, we utilize the high-performance computing mode in the CUDA environment for training and processing data, and the real-time performance has been improved significantly. The new model has been named: new GaitBase. In the study, we conducted experiments on multiple common gait recognition datasets, such as CASIA-B, GREW, and Gait3D. The results reflect that New GaitBase achieved significant results beyond previous gait recognition models (2023CVPR: GaitBase) under all conditions. The method achieves a significant improvement in Rank-1 accuracy compared to existing benchmarks: it outperforms CASIA-B by 2.4%, GREW by 29.1%, and Gait3D by 26.4%. Code is available at: https://github.com/yigemeinianda/New_GaitBase_version1.0.