Scholarlink

<p>In security authentication scenarios where faces may be obscured, gait serves as a reliable biometric cue for human identification. This paper presents RIGID, a fast and lightweight framework that relies solely on motion cues extracted from videos. Unlike silhouette-based methods, it directly exploits skeleton key points to achieve real-time recognition. The core novelty of RIGID is twofold. First, we introduce a tri-channel composite representation that jointly encodes coordinate maps, inter-joint distance matrices, and grayscale frames within a unified input tensor, thereby preserving both fine-grained skeletal geometry and complementary appearance context. Second, to the best of our knowledge, we present a novel adaptation of the ConvNeXtV2 architecture to gait recognition, enabling attention-free and computationally efficient modeling of discriminative motion patterns. Built upon this design, RIGID achieves 96.42% accuracy with only 578K parameters and 3.02 ms/frame inference on a recent public benchmark of 64 subjects and 3,120 videos captured in diverse indoor and outdoor settings. Additional evaluation on our collected RealGait-v2 dataset, featuring occlusion, low-light, and clothing variation, provides indicative evidence of robustness in real-world conditions. Overall, RIGID offers a balanced solution that advances the practical deployment of gait recognition through joint optimization of accuracy, efficiency, and environmental robustness.</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

RIGID: real-time indexing of humans via gait identification and detection

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • Hoang-Tuan Dao-Xuan,
  • Khanh-Duy Cao-Phan,
  • Van-Linh Nguyen

摘要

In security authentication scenarios where faces may be obscured, gait serves as a reliable biometric cue for human identification. This paper presents RIGID, a fast and lightweight framework that relies solely on motion cues extracted from videos. Unlike silhouette-based methods, it directly exploits skeleton key points to achieve real-time recognition. The core novelty of RIGID is twofold. First, we introduce a tri-channel composite representation that jointly encodes coordinate maps, inter-joint distance matrices, and grayscale frames within a unified input tensor, thereby preserving both fine-grained skeletal geometry and complementary appearance context. Second, to the best of our knowledge, we present a novel adaptation of the ConvNeXtV2 architecture to gait recognition, enabling attention-free and computationally efficient modeling of discriminative motion patterns. Built upon this design, RIGID achieves 96.42% accuracy with only 578K parameters and 3.02 ms/frame inference on a recent public benchmark of 64 subjects and 3,120 videos captured in diverse indoor and outdoor settings. Additional evaluation on our collected RealGait-v2 dataset, featuring occlusion, low-light, and clothing variation, provides indicative evidence of robustness in real-world conditions. Overall, RIGID offers a balanced solution that advances the practical deployment of gait recognition through joint optimization of accuracy, efficiency, and environmental robustness.