Scholarlink

<p>Gait recognition is a non-contact biometric modality with increasing relevance in surveillance and healthcare applications. This work presents a computationally efficient identity recognition framework based on a customized MobileNet architecture tailored for silhouette-based gait analysis. Unlike heavy CNN backbones traditionally used for gait recognition, the proposed approach adapts depthwise separable convolutions and Inverted Residual Blocks (IRBs) for lightweight deployment while preserving discriminative spatial representations. The model is evaluated on the CASIA-A dataset under multiple training protocols and demonstrates competitive accuracy (up to 95.66%) with substantially reduced inference time (12&#xa0;ms per image) and memory consumption (28&#xa0;MB) compared to VGG19 (41&#xa0;ms, 145&#xa0;MB). The study further investigates hyperparameter sensitivity, resolution trade-offs, and ablation analysis to quantify the contribution of architectural components. The results indicate that carefully optimized lightweight CNNs can provide a favorable balance between recognition performance and computational efficiency for real-time gait-based biometric systems.</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

Deep Learning-Based Gait Recognition with Optimized Hyperparameters Using MobileNet

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Sakinah Mohd Shukri,
  • Mohammed Yousif Abo Keir,
  • Amit Sandhu,
  • G. Sridevi,
  • R. Manjunatha,
  • Nilesh Bhosle,
  • Debasish Shit,
  • Ahmed Alkhayyat

摘要

Gait recognition is a non-contact biometric modality with increasing relevance in surveillance and healthcare applications. This work presents a computationally efficient identity recognition framework based on a customized MobileNet architecture tailored for silhouette-based gait analysis. Unlike heavy CNN backbones traditionally used for gait recognition, the proposed approach adapts depthwise separable convolutions and Inverted Residual Blocks (IRBs) for lightweight deployment while preserving discriminative spatial representations. The model is evaluated on the CASIA-A dataset under multiple training protocols and demonstrates competitive accuracy (up to 95.66%) with substantially reduced inference time (12 ms per image) and memory consumption (28 MB) compared to VGG19 (41 ms, 145 MB). The study further investigates hyperparameter sensitivity, resolution trade-offs, and ablation analysis to quantify the contribution of architectural components. The results indicate that carefully optimized lightweight CNNs can provide a favorable balance between recognition performance and computational efficiency for real-time gait-based biometric systems.