Combining Multiscale Feature Fusion with Frame Attention Mechanism for Emotion Recognition from Body Movements

Aiming at the problems such as the different scales information and low recognition accuracy of emotion recognition from body movements using Convolutional 3D (C3D) network. We propose emotion recognition model from body movements that combines multiscale feature fusion and frame attention mechanism. Firstly, we use the C3D network as the base and construct the C3D-5 network with a convolutional kernel size of 3x5x5. Batch Normalization (BN) layer is added to both the C3D and C3D-5 networks. Then we fuse the output features of the C3D and C3D-5 networks to capture body movement information across different scales. Finally, we introduce a Frame Attention (FA) mechanism, applying BN layers to normalize multiscale features, enabling the model to focus on key feature information across different scales during feature fusion. The experimental results show that the model achieves an emotion recognition accuracy of 82.3% from body movements on the FABO dataset, confirming its effectiveness for emotion recognition.