Design And Implementation Of Online Multi-person Action Recognition System Based On Sequence Frame Analysis

In recent years the demand for intelligent monitoring in the security field is growing,with online multi-person action recognition function of the system can be timely detection of dangerous behavior so as to prevent problems before they occur,and therefore gradually by the industry’s attention.In the entertainment industry,vision-based action recognition algorithms can also greatly reduce the cost of equipment for physical games.At the same time,action recognition also brings a new model for human-computer interaction.Video action recognition algorithms can currently be divided into two branches: traditional methods based on manual feature extraction and deep learning.Early traditional visionbased methods suffer from the problem of insufficient representation of action information by artificially designed features and low accuracy,while deep learning-based methods generally have difficult model training and slow recognition speed.In this paper,we combine traditional methods with deep learning,firstly,we improve and optimize Openpose to form the front skeleton output module of the action recognition system in this paper,then we propose a unique artificial feature extraction method for the skeleton sequence data output from this module to obtain the mixed-order feature representation of action sequences,and finally,we input these feature data into a special action recognition network based on multibranching one-dimensional convolutional Finally,these feature data are fed into a special action recognition network based on multi-branching one-dimensional convolutional design for fast action classification,thus constituting a complete online multiplayer action recognition system,which guarantees the prediction accuracy while improving the operation speed and achieving the effect of real-time prediction.the main work is as follows:(1)According to the action analysis,11 main key points of the human body are extracted to form a human skeleton composed of 10 skeletons.Compared with openpose,the skeleton in this paper removes the face and other key points that do not contribute much to the pose estimation,and grasps the human body.The main features of the action are reduced while the amount of computation is reduced.(2)The NTU+RGB+D data set used in the training model in this paper is processed,the invalid data in the data set is removed,the multi-person interaction actions in the data set are also removed,and 49 single-person action categories are retained.Since the number of frames of each action sample in the native data set is uncertain,this paper adopts uniform sampling to set the length of all data samples to 50 frames for subsequent training models.(3)A simplified skeleton extraction network is proposed based on Openpose.This network can output the feature skeleton defined in(1),and at the same time,compared with the original Openpose,the calculation time is greatly reduced.(4)A fusion one-dimensional convolutional network based on skeleton frame input is proposed.This network performs one-dimensional convolution from three aspects: angle sequence,angular velocity sequence,and angular acceleration sequence,and then performs feature fusion.Finally,the output layer is classified and output.The network structure in this paper is the optimal solution obtained through experimental tuning,which has the characteristics of fast convergence and high accuracy.(5)For the multi-target tracking problem caused by the presence of multiple people in the video,this paper proposes an affinity coding based on skeleton information,and then applies the KM algorithm for multi-person skeleton tracking on this basis,and finally integrates the system modules into a complete online multi-person action recognition system.