Gesture estimation for 3D martial arts based on neural network

Abstract

A 3D martial arts gestures estimation method is proposed, for it is difficult to accurately describe the gestures, which is caused by the high degree of freedom and high similarity of 3D martial arts gestures. This method is combined with finger kinematics analysis model. And beyond that, it is based on the morphological topological structure of hand and combined with the CNN neural network. Firstly, the CNN is used to extract and classify 3D gesture features of martial arts. Then, using the morphological topological structure of hands to simulate the dependence of hand joints, and the 3D coordinates of hand joints were obtained. Finally, the attitude regression module is used to realize the attitude estimation of martial arts gesture action. Simulation results show that the accuracy of gesture estimation can be improved through cascade splicing in this research. Compared with existing 3D gesture estimation methods such as V2V, Pose-REN and CrossInfoNet, the proposed method performs better in MSE and FS indicators. It has lower estimation errors and an inference speed of 220.7 frames per second. In addition, the three-dimensional visualization results show that the predicted joint points obtained by the proposed estimation method coincide with the labeled joint points, and there is no occlusion. So it is proved that the proposed method is feasible, and it can be used to estimate martial arts gestures in the future.