3D Skeleton-based Human Action Recognition Algorithm Research

With the advent of the intelligent era,human action recognition which act as a key technology for many practical applications and needs has become one of the most active research directions in computer vision.Early study mainly extracted spatio-temporal features from RGB videos captured by 2D cameras.With the emergence of the depth camera,human action recognition can seek a solution in three-dimensional space,that is,action recognition based on 3D skeleton.Compared with RGB videos,3D skeleton data are insensitive to lighting,background and occlusion,and has a smaller data dimension,which are advantage of it.Therefore,3D skeleton-based action recognition has important research value.In recent years,the method for 3d skeleton-action recognition exploiting deep convolutional neural network emerged endlessly and have gained breakthrough progress.However,how to use the convolutional neural network to utilize the topological structure of skeleton data,how to fully explore the potential associations implied in multiple features of skeleton data and how to strengthen the robustness of view-point changes need to be further studied.In this paper,we will research from three aspects which are multiple feature,multiple view-point and potential association of multiple feature.The main contributions of this paper are listed as follows:(1)This paper designs the bones feature according to the topological structure of human body.And combine the bones with skeleton positions and motions together to represent skeletal video.(2)Based on the Euler angle rotation formula,this paper proposes the multiple view-point transformer,which improves the robustness of changes in view-point by simulating and synthesizing action information in multiple perspective.(3)This paper proposes the multiple features pairwise fusion to take full advantage of the potential associations between multiple features.Meanwhile,this paper proposes the pattern of multi-task learning combined with ensemble learning to train model and ensemble result in order to further enhance the generalization ability of model.Finally,this paper proposes the three-stream convolutional neural network(3SCNN)based on the above components,and performs comparative experiments and ablation experiments on mainstream skeleton-based human behavior dataset,which fully proves the effectiveness of our method.