Dynamic Gesture Recognition Based on Deep 3D Natural Networks

The process of establishing contact with a computer through gestures still faces many bottlenecks that need to be solved, one of which is how to accurately recognize dynamic gestures. The traditional 3D convolution operation is transformed into a separable convolution operation based on separable convolution. The larger convolution kernel is transformed into smaller convolution and point convolution operations in the depth direction of the input data. The purpose of extracting and modeling the features of the input data is achieved, as well as the purpose of reducing the amount of calculation for the 3D convolution operation. Therefore, a long and short time memory network (which is more sensitive to temporal features) is introduced in the underlying network structure to memorize and model the relevant temporal information. Additionally, in order to overcome the adverse effects of the external background environment, an attention mechanism is also introduced into the structure to achieve suppression of the irrelevant background. On the basis of the existing research, by fusing the information of different modalities, the designed related experiments verify that the fusion operation is better than the recognition effect of single-modal information in dynamic gesture tasks. The experimental results show that three network structures, the efficient extraction of dynamic gesture features, and better recognition performance on the validation set of the IsoGD dataset can all be achieved. We propose a new method for dynamic gesture recognition, and the proposed model effectively improves the accuracy of gesture recognition.