A Dynamic Head Gesture Recognition Method for Real-Time Human-Computer Interaction

In human-computer interaction, head gestures play a significant role in improving smoothness and naturality. However, existing head gesture recognition algorithms have disadvantages in accuracy and generalization ability. To deal with these problems, this paper addresses a two-stream dynamic head gesture recognition method with the SlowFast pathway called 3DSFI (3D SlowFast Inception). The SlowFast pathway is designed to reduce parameters and computational costs. Meanwhile, its two-stream structure can efficiently capture motion features in videos and dense optical flows. Besides, Inception blocks of InceptionV3 are expanded by the 3D convolutional kernel into space-time and serve as a feature extractor. Finally, 3DSFI is applied to a robot Pepper in order to evaluate realistic performance. Experimental results show that the proposed method has higher accuracy and better generalization performance than the classical C3D (Convolutional 3D) and I3D (Inflated 3D ConvNet) methods.