Research And Implementation On Multi-Person Activity Recognition Technology Based On Pose Estimation

In rapidly developing information era,human activity recognition(HAR)technology based on computer vision is widely used in the fields of smart home,smart medical treatment,and virtual reality,etc.Especially in recent years,with the development of deep learning technologies,such as convolutional neural networks(CNN)and recurrent neural networks(RNN),the HAR technology based on computer vision has made many breakthroughs in feature automatic extraction and recognition accuracy,showing a great application prospectAlthough the HAR technology based on computer vision has made great progress,it still faces many problems and challenges.For example,most of the current HAR research is single-person HAR research,and there are few researches on multi-person HAR.In actual life,multi-person activity or group activity accounts for a large proportion Although the existing multi-person HAR system framework can complete the recognition task,the steps are cumbersome,the efficiency is low,and it takes a long time in running.In addition,there are still many areas that need to be improved in terms of multi-person detection,external feature extraction,and recognition accuracy.In order to solve the above problems,this thesis has conducted in-depth research on system framework optimization,feature extraction method,classification recognition and visual display of results,mainly completing the following aspects:(1)Research on multi-person activity recognition system and skeleton keypoint extraction method.For the problems that the current multi-person HAR systems are imperfect and the number of people in the image is difficult to determine,a multi-person HAR system based on skeleton keypoint detection is proposed.Meanwhile,a skeleton keypoint extraction method based on OpenPose pose estimation model is proposed.(2)Research on multi-person posture feature extraction method.Combined with the characteristics of skeleton keypoint data,three effective features of human pose are extracted.At the same time,in order to accurately describe the relationship between features,a feature description method based on frame window matrix is proposed,Simulation experiments has been conducted on public datasets,and the experimental results show that the proposed method can effectively improve the accuracy of activity recognition,and has good robustness and stability.(3)Research on double classification recognition and design of visual display interface.For the proposed multi-person HAR system,in order to further improve the activity recognition accuracy,the support vector machine(SVM)classifier and the naive bayes(NB)network are combined to form a double classifier for multi-person activity recognition.Simulation experiments have been carried out on public datasets,the experimental results show that the proposed method can further improve the activity recognition accuracy.At the same time,in order to display the recognition results of the proposed system more clearly,a visual interface is designed based on the Python language to display the final recognition results.This thesis mainly uses the UT-Interaction and HMDB51 datasets as simulation inputs.The simulation results show that the skeleton keypoint extraction method based on OpenPose can well detect the number of people in the image and extract useful human skeleton keypoints.The feature description method based on the frame_window matrix can effectively extract the posture characteristics of multiple people.The double classification method can further improve the activity recognition accuracy.The visual display interface can clearly show the results of the multi-person HAR system.And the research results of this thesis can be applied to practical fields such as smart classroom,modern prison,virtual reality,etc.,which have certain practical significance and social value.