Intercity rail platform abnormal action recognition based on a skeleton tracking and recognition framework

The skeleton data of limbs are unreliable due to occlusion and camera viewpoints on intercity railway platforms. It hinders the acquisition of skeleton sequences and disturbs the skeleton-based abnormal action recognition. To overcome these issues, this work proposes a framework consisting of a pose tracking module and an abnormal action recognition module. The proposed pose tracking module maintains the identities of multiple human poses across frames and provides skeleton sequences as input for recognition. Instead of utilizing the whole skeleton, the pose tracking method tracks the trunk for more stable results of identity association as the estimations of the limbs are unreliable. In addition, a position embedding graph convolutional network (PEGCN) is proposed to recognize abnormal actions. PEGCN utilizes a simple cosine encoding as position embeddings for enhancing the differentiation of skeleton vertices and an SElayer for extracting temporal dynamics. The pose tracking method achieves 66.42% tracking accuracy scores and higher frame rates than previous methods on the PoseTrack dataset. Additionally, PEGCN achieves competitive results on the Intercity Railway Action Dataset (IRAD) and the public NTU-RGB+D dataset.