Research On Action Analysis Based On Human Skeleton Information

Human motion analysis has always been a research hotspot in the field of computer vision,Human motion recognition is an important branch of it.In recent years,with the development of big data and graphics processors,RGB video-based motion recognition algorithms have achieved rapid development.It has a wide range of applications in the fields of intelligent monitoring,medical care and human-computer interaction.With the development of depth cameras,the acquisition of human skeleton data has become easier.Compared with RGB video,human skeleton joint points can accurately describe the process of human movement,and can solve the problem that RGB video is sensitive to light and is greatly affected by occlusion.People began to focus more on skeleton-based action recognition research.The core of skeleton-based action recognition lies in the selection of feature extraction and classification methods.This paper develops research based on skeleton feature extraction and proposes a feature extraction method.At the same time,the accuracy of the action is described on this basis.The main work content is as follows:(1)Aiming at the problem of feature extraction based on skeleton action recognition,a feature modeling method based on Point-Line is proposed.First select32 important lines between the 16 main joint points as the bottom edge,extract the distance from each joint point(Point)to 32 bottom edges(Line)as features,and combine the feature representation method of tree traversal to perform feature modeling;The Attention of Context Fixed Long Short Term Memmory(AF-LSTM)method is proposed for classification,and the Point-Line feature is used as the input of the AF-LSTM network to complete the action recognition task.The proposed method is tested on NTU RGB+D,and the experiment proves that the method proposed in this paper improves 5.51% and 7.10% on C-S and C-V respectively compared with the method without Point-Line modeling.(2)In order to describe the accuracy of actions of the same category relative to standard actions,an accuracy evaluation mechanism is proposed.Firstly,the improvement strategy of dynamic time warping algorithm(DTW)is integrated,and a Fixed matching algorithm(FDTW)is proposed.FDTW is used to calculate the similarity between user action sequence and standard action sequence,aiming at different lengths and different sizes.Set the accuracy evaluation criteria for the time series respectively,and finally use the Fourier transform to fit the FDTW value to the final percentile score.We carried out test experiments on the accuracy and running time of FDTW on Tai Chi movement data.Experiments show that in terms of accuracy verification,when the user sequence is similar to the standard sequence,the average error rate is 9.82%.In the time test,the running time is greatly shortened.(3)Based on the Unity 3D platform,a real-time Tai Chi action guidance platform was built.First,use the action recognition method of this article to detect the user's gestures,select different modes through different gestures,and then use Unity's particle system to build virtual special effects to realize the action guidance function,and finally use the accuracy evaluation mechanism of this article to score the user's Tai Chi movements.