Real-time gesture recognition in complex background based on deep learning
Helin Long, Muyang Liu, Muzhi Li · 2020
The non-contact gesture recognition system is essential to enhance the immersive experience of human-computer interaction. In recent years, the gesture recognition network is difficult to deal with complex backgrounds and the diversity of shooting angles, which limits its scope of application. This paper proposes a two Staged gesture recognition network: firstly locate the gesture through the detection network, and then use the classifier to identify the gesture category in the localized area. Through BiFPN for efficient multi-scale semantic fusion, the network can reach 96.73% gesture recognition accuracy in complex background. At the same time, the network is further compressed through pruning and weighting, so that the trained network can be deployed on embedded terminals to realize real-time gesture recognition.