Research And Development Of MSFAFN Fitness Action Analysis System Based On LSTM

With the improvement of living standards and the popularization of health concepts,people’s awareness of fitness exercise has become stronger and stronger.In the context of normalized prevention and control of the new crown epidemic,people’s demand for home fitness is constantly increasing.One of the modes of home fitness is to use the online fitness action analysis system for online fitness.Since the existing system can only identify the type of action,it cannot analyze the specific action,and cannot give the user detailed action correction suggestions.The basis of the system is the human body motion pose estimation network model,and the existing model has the problems of large scale and slow processing speed,which leads to the long response time of the system.In view of the above problems,this paper studies and implements a mobile terminal-based fitness process action analysis system,including:Aiming at the large scale and slow speed of the existing human pose estimation network model,this paper constructs a lightweight human pose estimation network model Shuffle Pose.The model is based on Open Pose.First,Shuffle Net V2 is introduced as the backbone network to improve the attitude estimation speed.Second,the sequential network structure is used to reduce the network width.Finally,the residual network and small convolution kernel are used to slow down the network degradation and reduce the network width network size.The experimental results show that,compared with Open Pose,the FPS index of this model is increased by nearly 50%,and the detection accuracy is increased by 1.8%;compared with the existing lightweight pose estimation model Efficient Pose,the FPS index is 2 frames/second higher.In order to describe the body movements in the fitness process in detail,based on the coordinate sequence of skeletal key points obtained by Shuffle Pose,this paper converts the movement analysis in the fitness process into a multi-label classification problem based on time series,and constructs a multi-scale feature attention based on LSTM.Force Fusion Network(MSFAFN-LSTM).First,the LSTM branch is introduced to schedule the transmission of sequence information;secondly,the MSFAFM branch is used to extract multiscale features,and the channel attention mechanism is used to further obtain the time series features of key regions to obtain multi-scale feature information of important regions.The experimental results show that the MSFAFN-LSTM model is 1% higher than the existing models in the F1 indicator.Finally,the correlation of labels is analyzed by the experimental results to simplify the classification process.Through experimental comparison,the simplified MSFAFN-LSTM classification accuracy is improved by 7%,and the processing time is reduced by 12 seconds.Based on the constructed network model,this paper designs and implements a fitness action analysis system with front-end and back-end separation.First,the front-end interacts with users through the Android App client;secondly,the back-end uses distributed deployment and micro-service architecture,and uses message queues to realize information transmission between services.Finally,the system shows the user’s problems and suggestions for improvement during the fitness process,which further verifies the effectiveness and applicability of the system.