Scholarlink

<p>Numerous biometric techniques are currently available for identifying humans. Each of the present techniques has some drawbacks that make them vulnerable, such as facial recognition systems becoming obsolete during face transplant surgery. Gait identification is a biometric recognition technique that distinguishes individuals based on their unique gait patterns. It is used in various fields such as clinical analysis, action recognition, security, surveillance, and monitoring. Gait recognition is unique to individuals due to anatomical, physiological, and behavioral factors that influence their walking patterns. These factors create different variations in stride length, cadence, posture, and other gait characteristics. Gait recognition uses these unique features to accurately differentiate and identify individuals. Hence, it is almost impossible to mimic the gait of any individual compared to any other biometric identification technique. This biometric method is non-invasive, meaning it does not require any physical contact or special participation, making it ideal for surveillance and remote identification. In this work, we have developed a gait identification model by projecting the Silhouettes as a structure of nodes and edges, which enhances the Silhouette’s spectral features and helps identify the individuals. The CASIA-B data set is used to train the model, consisting of 121 silhouettes of people from different angles. Sequences of 12 silhouettes are used to extract the Gait Energy Images (GEIs), and the extracted GEIs are projected as the structures of the nodes and edges. This is treated as a dataset to train the Deepwalk as a feature extractor and with a 1D-CNN-LSTM-based classification model. The trained model outperformed the existing models as it achieved an accuracy of 0.9978, a recall of 0.9923, a precision of 0.9993, and an F1-score of 0.9957 in gait identification. In addition, we conducted the ablation study in order to confirm the importance of the use of LSTM and 1D-CNN models. Compared to existing models, which used CASIA-B data, the proposed model generated better accuracy.</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

Gait identification based on deepwalk features using CNN and LSTM: an advanced biometric approach

  • 期刊论文

  • 发表日期：

    2025年6月9日

查看全文
  • Ravi Shekhar Tiwari,
  • Tapan Kumar Das,
  • Asis Kumar Tripathy,
  • Kuan-Ching Li

摘要

Numerous biometric techniques are currently available for identifying humans. Each of the present techniques has some drawbacks that make them vulnerable, such as facial recognition systems becoming obsolete during face transplant surgery. Gait identification is a biometric recognition technique that distinguishes individuals based on their unique gait patterns. It is used in various fields such as clinical analysis, action recognition, security, surveillance, and monitoring. Gait recognition is unique to individuals due to anatomical, physiological, and behavioral factors that influence their walking patterns. These factors create different variations in stride length, cadence, posture, and other gait characteristics. Gait recognition uses these unique features to accurately differentiate and identify individuals. Hence, it is almost impossible to mimic the gait of any individual compared to any other biometric identification technique. This biometric method is non-invasive, meaning it does not require any physical contact or special participation, making it ideal for surveillance and remote identification. In this work, we have developed a gait identification model by projecting the Silhouettes as a structure of nodes and edges, which enhances the Silhouette’s spectral features and helps identify the individuals. The CASIA-B data set is used to train the model, consisting of 121 silhouettes of people from different angles. Sequences of 12 silhouettes are used to extract the Gait Energy Images (GEIs), and the extracted GEIs are projected as the structures of the nodes and edges. This is treated as a dataset to train the Deepwalk as a feature extractor and with a 1D-CNN-LSTM-based classification model. The trained model outperformed the existing models as it achieved an accuracy of 0.9978, a recall of 0.9923, a precision of 0.9993, and an F1-score of 0.9957 in gait identification. In addition, we conducted the ablation study in order to confirm the importance of the use of LSTM and 1D-CNN models. Compared to existing models, which used CASIA-B data, the proposed model generated better accuracy.