Research And Application Of Human Action Recognition Based On Skeleton Sequence

Equality and justice are important parts of socialist core values,preventing and eliminating cheating in examinations is an important measure to safeguard the fairness of education and employment.In recent years,through improving the examination site management system,improving the professional ability of the supervision team,increasing the video monitoring coverage of the examination room,the education and employment equity related to the examination have been significantly improved.But in the unified entrance examination of national graduate students,the examiners should keep their attention in the last four hours.They should not only prevent cheating,but also pay attention to the physiological and psychological state of the examinees.The examiners have a large workload and high psychological pressure.After the examination,a large number of manpower should be organized to review the surveillance videos of the examination one by one,which is extremely inefficient.With the development of computer vision and artificial intelligence technology,it is possible to automatically monitor the dynamic of the examination and identify the abnormal action of the examination through video analysis.This paper studies the method of abnormal action recognition of examinees based on video,and designs and implements an intelligent video examination assistant system.The system can perceive and identify the common cheating actions,abnormal actions and states of examinees,timely feedback to the examiners and examination managers,and collect the evidences and records of cheating actions.It is not only conducive to ensuring the order of the examination and maintaining the fairness of the examination,find and cure the examinees of sudden diseases in time,but also to relieve the workload and psychological pressure of the examiners,and improve the quality and efficiency of the examination management.The main research contents of this paper include the following three aspects:(1)A spatio-temporal adaptive graph convolution action recognition method based on motion information is proposed.With the development and maturity of human pose-estimation methods,combining human posture and graph convolution network has become an important research direction of human action recognition.This method can express the connection relationship between joints fully by representing the structure of the joints as a graph.However,it is difficult to recognize and make full use of the key frames and key joints characteristics in the action sequence by end-to-end learning.In order to solve this problem,this paper captures the motion information between frames,and adaptively enhances the features of key frames and key joints in the process of graph convolution according to the intensity of motion.The experimental results show that the method is simple and effective.(2)A new method of action recognition based on spatio-temporal adaptive and high resolution graph convolution network is proposed.The common results of human pose-estimation only include 18 key points,and 18 key points may not fully express the spatial information of human posture.Therefore,two schemes are proposed to build virtual key points to improve the resolution of human skeleton.In addition,a fully connected graph convolution layer is added between the last layer of the general spatio-temporal graph convolution and the global average pool layer in the graph convolution network model.The action recognition can be performed by using the correlation information between the long-distance key points without significantly increasing the computational complexity.Combined with the spatio-temporal adaptive graph convolution,a new spatio-temporal adaptive high resolution graph convolution network based on motion information is finally realized.The experimental results on NTU-RGB+D and Kinetic-Skeleton data sets show that the proposed network action recognition accuracy is better than the existing method.(3)The design and implementation of intelligent video examination assistant system.By using the proposed action recognition method and mature human detection and pose-estimation methods,this paper designs and implements an intelligent video examination assistant system.The system can perceive and identify common cheating actions,abnormal actions and states of examinees,timely feedback to the examiners and examination managers,and collect evidences and records of cheating actions.This paper uses python3.6 language,Py Qt5 toolkit and pytoch1.0 deep learning framework to complete the development in Ubuntu operating system.The system runs smoothly,the interface is simple and friendly,and has certain application value.