Liu Jinhang, Ke Yunfan, Zhou Ting, Qiu Yan, Wang Chunzhi
School of Computer Science, Hubei University of Technology, Wuhan 430068, China.
Department of Computer Science, College of Engineering and Technology, Hubei University of Technology, Wuhan 430068, China.
Sensors (Basel). 2025 Apr 8;25(8):2337. doi: 10.3390/s25082337.
Gait recognition, a long-range biometric technique based on walking posture, the fact that they do not require the cooperation of the subject and are non-invasive has made them highly sought after in recent years.Although existing methods have achieved impressive results in laboratory environments, the recognition performance is still deficient in real-world applications, especially when confronted with complex and dynamic scenarios. The major challenges in gait recognition include changes in viewing angle, occlusion, clothing changes, and significant differences in gait characteristics under different walking conditions. To slove these issues, we propose a gait recognition method based on relational-aware global attention. Specifically, we introduce a Relational-aware Global Attention (RGA) module, which captures global structural information within gait sequences to enable more precise attention learning. Unlike traditional gait recognition methods that rely solely on local convolutions, we stack pairwise associations between each feature position in the gait silhouette and all other feature positions, along with the features themselves, using a shallow convolutional model to learn attention. This approach is particularly effective in gait recognition due to the physical constraints on human walking postures, allowing the structural information embedded in the global relationships to aid in inferring the semantics and focus areas of various body parts, thereby improving the differentiation of gait features across individuals. Our experimental results on multiple datasets (Grew, Gait3D, SUSTech1k) demonstrate that GaitRGA achieves significant performance improvements, especially in real-world scenarios.
步态识别是一种基于行走姿势的远距离生物识别技术，由于其无需被试者配合且具有非侵入性，近年来备受关注。尽管现有方法在实验室环境中取得了令人瞩目的成果，但在实际应用中识别性能仍存在不足，尤其是在面对复杂动态场景时。步态识别中的主要挑战包括视角变化、遮挡、服装变化以及不同行走条件下步态特征的显著差异。为了解决这些问题，我们提出了一种基于关系感知全局注意力的步态识别方法。具体而言，我们引入了关系感知全局注意力（RGA）模块，该模块捕捉步态序列中的全局结构信息，以实现更精确的注意力学习。与传统仅依赖局部卷积的步态识别方法不同，我们使用浅层卷积模型，将步态轮廓中每个特征位置与所有其他特征位置之间的成对关联以及特征本身进行堆叠，以学习注意力。由于人类行走姿势的物理限制，这种方法在步态识别中特别有效，它能让嵌入在全局关系中的结构信息有助于推断各个身体部位的语义和重点区域，从而提高不同个体之间步态特征的区分度。我们在多个数据集（Grew、Gait3D、SUSTech1k）上的实验结果表明，GaitRGA取得了显著的性能提升，尤其是在实际场景中。
