A Feature Fusion Network for Skeleton-Based Gesture Recognition

With the development of the times, the requirements for human-computer interaction methods have gradually increased, and naturalness and comfort are constantly pursued on the basis of traditional precision. Gesture is one of the innate ways of human communication, which is highly intuitive and can be employed as an effective means of natural human-computer interaction. In this paper, dynamic gestures are investigated based on the 3D skeletal information of gestures, and different cropping boxes are placed at generate global and local datasets respectively according to whether they depend on the motion trajectory of the gesture. By analyzing the geometric features of skeletal sequences, a dual-stream 3D CNN (Double_C3D) framework is proposed for fusion at the feature level, which relies on 3D heat map video streams and uses the video streams as the input to the network. Finally, the Double_C3D framework was evaluated on the SHREC dynamic gesture recognition dataset and the JHMBD dynamic behavior recognition dataset with an accuracy of 91.72% and 70.54%, respectively.