Gesture Recognition Based on Video Sequence with B3D Convolutional Neural Network

Gesture recognition has attracted great attention owing to its wide applications. However, until now, the current gesture recognition methods still have some drawbacks like the background handicap the improvement of recognition accuracy. in order to solve the issue, this paper presents a method based on a threedimensional convolution neural network and bi-directional LSTM networks, Which is named as B3D network. In this paper, Faster R-CNN is used to capture the information of hand position. The method of B3D Network is to capture the spatiotemporal feature informmation for a video representation and analyze temporal feature sequences. The experiment is conducted on DEVISIGN_D dataset and the lab self-built SGR dataset, Compared with other traditional methods such as Alex, C3D+LSTM and so on, Experiments results show that the object location strategy and bi-directional LSTM improved recognition accuracy, achieves 88.7% accuracy of SGR-Dataset, and 86.1% accuracy on DEVISIGN_D dataset, The recognition accuracy of our method both outperform traditional method.