Attention mechanism based multimodal feature fusion network for human action recognition

Current human action recognition (HAR) methods focus on integrating multiple data modalities, such as skeleton data and RGB data. However, they struggle to exploit motion correlation information in skeleton data and rely on spatial representations from RGB modalities. This paper proposes a novel Attention-based Multimodal Feature Integration Network (AMFI-Net) designed to enhance modal fusion and improve recognition accuracy. First, RGB and skeleton data undergo multi-level preprocessing to obtain differential movement representations, which are then input into a heterogeneous network for separate multimodal feature extraction. Next, an adaptive fusion strategy is employed to enhance the integration of these multimodal features. Finally, the network assesses the confidence level of weighted skeleton information to determine the extent and type of appearance information to be used in the final feature integration. Experiments conducted on the NTU-RGB + D dataset demonstrate that the proposed method is feasible, leading to significant improvements in human action recognition accuracy.