Temporal multiscale feature network for skeleton-based action recognition

Skeleton-based action recognition utilizes human skeletal data to identify and analyze human behaviors , which is widely applied in various scenarios such as intelligent surveillance, virtual reality, and autonomous driving. Since human actions usually have different durations, recognizing skeleton-based actions accurately and efficiently remains a challenge. To address the issue of temporal scale, this paper introduces a temporal multi-scale feature network (TMSF-Net) designed to enhance the recognition of skeleton-based actions. Specifically, TMSF-Net introduces a multi-scale temporal convolution module (MSTCM) to flexibly adjust the temporal receptive field of the network, enabling it to focus more on action-related regions. Additionally, TMSF-Net incorporates a Global Filter Module (GFM) to enhance the interaction among joint points across spatial and temporal dimensions. and adapt to different action scenarios. The efficacy of the proposed approach is demonstrated through experimental validation on two public datasets dedicated to skeleton action recognition.