Parallel temporal feature selection based on improved attention mechanism for dynamic gesture recognition

Abstract

Dynamic gesture recognition has become a new type of interaction to meet the needs of daily interaction. It is the most natural, easy to operate, and intuitive, so it has a wide range of applications. The accuracy of gesture recognition depends on the ability to accurately learn the short-term and long-term spatiotemporal features of gestures. Our work is different from improving the performance of a single type of network with convnets-based models and recurrent neural network-based models or serial stacking of two heterogeneous networks, we proposed a fusion architecture that can simultaneously learn short-term and long-term spatiotemporal features of gestures, which combined convnets-based models and recurrent neural network-based models in parallel. At each stage of feature learning, the short-term and long-term spatiotemporal features of gestures are captured simultaneously, and the contribution of two heterogeneous networks to the classification results in spatial and channel axes that can be learned automatically by using the attention mechanism. The sequence and pooling operation of the channel attention module and spatial attention module are compared through experiments. And the proportion of short-term and long-term features of gestures on channel and spatial axes in each stage of feature learning is quantitatively analyzed, and the final model is determined according to the experimental results. The module can be used for end-to-end learning and the proposed method was validated on the EgoGesture, SKIG, and IsoGD datasets and got very competitive performance.