Abstract: The challenges of low accuracy and poor generalization performance in recognizing abnormal behaviors of elevator passengers were attributed to the lack of sufficiently diverse abnormal behavior data. To address this issue, a Dynamic Graph Convolutional Network-based Behavior data Augmentation (DGCN-BA) framework was proposed. Firstly, a Dynamic Graph Convolutional Network (DGCN) was constructed to capture the spatial relationships and motion correlations among different human joints in the behaviors of elevator passengers. These features were utilized to enhance pose data, generating richer and more realistic pose sequences. Secondly, the pose sequences were used to construct human actions in a virtual elevator scene, generating a large amount of abnormal behavior video data for elevator passengers. Finally, to validate the effectiveness of DGCN-BA, experiments were conducted on public datasets Human3.6M, 3DHP, and MuPoTS-3D, as well as on a self-constructed dataset. Experimental results show that, compared to other data augmentation methods (JMDA (Joint Mixing Data Augmentation) and DDPMs (Denoising Diffusion Probabilistic Models)), DGCN-BA reduces the average MPJPE on the Human3.6M dataset by 2.9mm and 1.5mm, respectively. It more effectively handles pose estimation tasks, generates diverse and realistic abnormal behavior data, and significantly improves the recognition of elevator passenger abnormal behaviors in video-based scenarios.

Key words: graph convolutional network；data augmentation；behavior recognition；pose estimation；elevator passenger
 ,

摘要： 目前，因缺乏足够多样化的异常行为数据，电梯乘客异常行为的识别方法存在准确率不高和泛化性能较差的问题。为解决这个问题，提出一种基于动态图卷积网络的行为数据增强框架(DGCN-BA)。首先，构建一种动态图卷积网络，用于捕捉电梯乘客行为中不同人体关节之间的空间关系和运动关联性。首先，通过利用这些特征进行姿势增强，获取更丰富和合理的姿势序列。其次，利用姿势序列在虚拟电梯场景中构建人物动作，生成大量的电梯乘客异常行为视频数据。最后，为验证DGCN-BA的有效性，分别在公开数据集Human3.6M、3DHP、MuPoTS-3D和自建数据集上进行验证。实验结果表明，相较于其他的数据增强方法（如JMDA(Joint Mixing Data Augmentation)、DDPMs(Denoising Diffusion Probabilistic Models)），DGCN-BA在Human3.6M数据集上的平均MPJPE分别降低了2.9mm和1.5mm，能够更有效地进行姿势估计任务，生成合理多样的异常行为数据，并且明显改善了基于视频的电梯乘客异常行为识别效果。

关键词: 图卷积网络, 数据增强, 行为识别, 姿势估计, 电梯乘客