MDT: A multiscale differencing transformer with sequence feature relationship mining for robust action recognition

Skeleton-based action recognition, which analyzes joint coordinates and bone connections to classify human actions, is important in understanding and analyzing human dynamic behaviors. However, actions in complex scenes have a high degree of similarity and variability, with the dynamic changes in human skeletons and subtle temporal variations in particular posing significant challenges to the accuracy and robustness of action recognition systems. To mitigate these challenges, we propose a novel multiscale differencing transformer (MDT) with sequence feature relationship mining for robust action recognition. MDT effectively mines inter-frame timing information and feature distribution differences across multiple scales, enabling a deeper understanding of the nuances between actions. Specifically, we first propose multiscale differential self-attention to handle the need for understanding action changes across multiple time scales, improving the capacity of the model to effectively capture the global and local dynamic features of actions. Then, we introduce a sequence feature relationship mining module to address complex data patterns in scenes that may span multiple sequences, exhibiting both similar and distinct characteristics. By utilizing coarse- and fine-grained sequence information, this module empowers the model to recognize intricate data patterns. On the NTU RGB+D 60 dataset, the proposed MDT model outperforms the recent STAR-Transformer by 1.6% on the Cross-Subject (CS) setting and 1.1% on the Cross-View (CV) setting, demonstrating its consistent effectiveness across different evaluation protocols.