A Dynamic Gesture Recognition Method Based on R(2+1)D-Transformer Network

Efficient spatial-temporal feature extraction from input video streams is crucial for dynamic gesture recognition. In the task of video classification, convolutional neural networks (CNNs) are widely used as feature extractors, while methods based on recurrent neural networks (RNNs) are commonly employed for sequence modeling. However, RNNs lack the ability to model global dependencies and have a limited attention span in the temporal dimension. This becomes a performance bottleneck for dynamic gestures that require sensitivity to temporal correlations. To address this issue, this paper proposes a dynamic gesture recognition model called R(2+1)D-Transformer. It is a Transformer-based approach that focuses on global modeling. Firstly, the R(2+1)D network is employed as a spatial-temporal feature extractor to capture the spatiotemporal information. Then, self-attention-based Transformer is used to map the spatiotemporal feature sequence to the semantic representation of gesture movements, considering both the temporal and spatial context. Finally, the gesture recognition results are obtained through an MLP classification head. Experimental results demonstrate the effectiveness and potential of the proposed R(2+1)D-Transformer model on two publicly available dynamic gesture datasets, IPN-Hand and NvGesture. The promising performance of the proposed approach provides valuable insights and reference for further research and applications in dynamic gesture recognition.