Dynamic gesture recognition using 3D central difference separable residual LSTM coordinate attention networks

Abstract

The area of human–computer interaction has generated considerable interest in dynamic gesture recognition. However, the intrinsic qualities of the gestures themselves, including their flexibility and spatial scale, as well as external factors such as lighting and background, have impeded the improvement of recognition accuracy. To address this, we present a novel end-to-end recognition network named 3D Central Difference Separable Residual Long Short-Term Memory (LSTM) Coordinate Attention (3D CRLCA) in this paper. The network is composed of three components: (1) 3D Central Difference Separable Convolution (3D CDSC), (2) a residual module to enhance the network’s capability to distinguish between categories, and (3) an LSTM-Coordinate Attention (LSTM-CA) module to direct the network’s attention to the gesture region and its temporal and spatial characteristics. Our experiments using the ChaLearn Large-scale Gesture Recognition Dataset (IsoGD) and IPN datasets demonstrate the effectiveness of our approach, surpassing other existing methods.