The problems of low accuracy and poor generalization performance in recognizing abnormal behaviors of elevator passengers are due to the lack of sufficient diverse abnormal behavior data. To address these issues， a Dynamic Graph Convolutional Network-based Behavior data Augmentation （DGCN-BA） method was proposed. Firstly， a dynamic graph convolutional network was constructed to capture spatial relationships and motion correlations among different human joints in the behaviors of elevator passengers. Secondly， these features were utilized to enhance pose data， thereby generating richer and more reasonable pose sequences. Finally， the pose sequences were used to construct human actions in a virtual elevator scene， and lot of abnormal behavior video data for elevator passengers were generated. To validate the effectiveness of DGCN-BA， experiments were conducted on public datasets Human3.6M， 3DHP， MuPoTS-3D， and a self-constructed dataset. Experimental results show that compared to data augmentation methods JMDA （Joint Mixing Data Augmentation） and DDPMs （Denoising Diffusion Probabilistic Models）， DGCN-BA reduces the Mean Per Joint Position Error （MPJPE） on the Human3.6M dataset by 2.9 mm and 1.5 mm， respectively. It can be seen that DGCN-BA can complete pose estimation tasks more effectively， generates diverse and reasonable abnormal behavior data， and improves the recognition effect of video-based elevator passenger abnormal behaviors significantly.