Abstract: In video recognition, the traditional 2D convolution neural networks are easy to lose the relevant feature information in time dimension, which leads to the reduction of recognition accuracy. This paper uses 3D convolutional neural network as a basic network framework with 3D convolution kernel to extract the temporal and spatial features of videos, at the same time, the integration of multiple 3D convolutional neural network models are proposed to recognize dynamic gesture. In order to improve the convergence speed of the model and the stability of training, the network is optimized by Batch Normalization (BN) technology to shorten the training time of the network. Experimental results show that the proposed method has a good recognition performance for dynamic gesture recognition, and the recognition accuracy reaches 98.06% in Sheffield Kinect Gesture (SKIG) data set. Solely compared with RGB information, depth information and traditional 2D CNN, the gesture recognition rate is higher, which verifies the feasibility and effectiveness of the proposed method.

Key words: 3D Convolutional Neural Network (3D CNN), optical flow, ensemble learning, deep learning, dynamic gesture recognition