Dynamic gesture recognition method based on multi-head attention and multi-scale cavity convolution

Since the appearance of depth sensors like Kinect, 3D hand joint coordinate research has expanded. Many studies create gesture recognition systems by calculating joint dependencies, but complex relationships are hard to capture, making features ineffective. This paper proposes an improved model using multi - head attention and multi - scale cavity convolution. The multi - head attention encodes spatial relationships and captures node dependencies. Multi - scale convolution layers capture time information. An attention - weighted loss function boosts the model's robustness. Experiments on the DHG dataset show that for 14 - classified gestures, the model has a 97.8% accuracy, 7% higher than others, and for 28 - classified ones, 92.1% accuracy, 5% higher.