Research On Skeleton-Based Action Recognition With Capsule Neural Network

The research of skeleton-based action recognition algorithm is a novel and challenging topic in the field of computer vision.It is widely used in video information retrieval,virtual reality,human-computer interaction and other hot fields,and has achieved remarkable results.At present,the neural network algorithm used in most skeleton data action recognition models based on deep learning is graph convolution neural network.These models usually focus on modeling the relationship between adjacent joints in natural state,and then classify the extracted features by softmax and other relevant classification functions.However,this method does not make full use of the information of skeleton joints,such as the length,direction and motion of bones,and this feature classification function may destroy the overall structural connection between joint points,making the neural network ignore the indirect connection between some non-adjacent joint points,resulting in the insufficient accuracy of the neural network.In view of the shortcomings of most current models,this paper designs an improved scheme based on the action recognition network algorithm architecture of human skeleton data and the characteristics of human skeleton in motion.By analyzing the interaction between human skeleton and joint points in the process of movement,this paper proposes to add skeleton information to the process of feature extraction and update,so that multiple joint points that are far apart but closely related in the same movement can quickly perceive the information of each other.This paper attempts to introduce the capsule network and dynamic routing into the skeleton-based action recognition network algorithm.According to the characteristics of human skeleton and joint points during movement,features are encapsulated into low-lever capsules,and EM dynamic routing are used to cluster multiple low-lever capsules into high-level capsules representing different actions to improve the accuracy of classification tasks.In this paper,the one of largest open-source human skeleton data set,NTU RGB+D is used as the data set.And large number of experiments are carried out to evaluate the proposed scheme,which verifies the effectiveness and feasibility of this method.The accuracy rate on the NTU RGB+D dataset is 90.33%,which is 0.47% higher than the benchmark model.