基于RGB图像的三维骨骼轨迹的步态识别方法

Gait Recognition Method Based on 3D Skeletal Trajectories from RGB Images

摘要: 步态识别作为一种无接触的人物身份识别方法，近年来受到广泛关注。传统的基于骨骼点的步态识别方法主要依赖二维图像特征，易受环境因素影响。针对这一问题，本文提出了一种结合RGB图像序列与深度信息的步态识别方法，以提高识别准确性和鲁棒性。首先，本文使用OpenPose算法从视频帧中提取人体关键骨骼点，并通过熵度量筛选出对步态识别贡献较大的关节。与传统方法不同，本文将每个骨骼点的二维坐标转换为三维空间坐标，增强步态特征的表达能力。随后，通过对三维轨迹建模，构建包含时间序列信息的骨骼点轨迹描述符，并通过计算骨骼点间的相对距离和位移变化，捕捉运动特征。此外，结合时空信息与深度数据，进一步丰富了骨骼点轨迹描述，构建了具有更高辨识能力的步态特征向量。为提升识别准确性，本文采用LCSS相似度度量方法，并结合KNN分类器对步态序列进行分类，实现人物身份识别。

Abstract: Gait recognition, as a non-contact method for human identity recognition, has attracted widespread attention in recent years. Traditional skeleton-based gait recognition methods mainly rely on 2D image features, which are easily affected by environmental factors. To address this issue, this paper proposes a gait recognition method that combines RGB image sequences with depth information to improve recognition accuracy and robustness. First, the OpenPose algorithm is used to extract human key skeleton points from video frames, and an entropy-based measure is applied to select the joints that contribute more significantly to gait recognition. Unlike traditional methods, this paper converts the 2D coordinates of each skeleton point into 3D spatial coordinates to enhance the expressive power of gait features. Subsequently, by modeling the 3D trajectories, a skeleton point trajectory descriptor containing temporal sequence information is constructed, capturing motion characteristics by calculating the relative distance and displacement changes between skeleton points. Furthermore, by incorporating spatiotemporal information and depth data, the skeleton point trajectory description is enriched, resulting in a gait feature vector with higher discriminative ability. To improve recognition accuracy, the LCSS similarity measure is used, and the KNN classifier is applied to classify the gait sequences, achieving human identity recognition.