Research On Key Algorithms Of Soldier Physical Training Action Evaluation Based On Skeleton Key Point Detection

Physical training as one of the important means to improve the combat effectiveness of the troops,the standardization of movements directly affects the training effect.In order to ensure effective physical training,assessment and guidance of training movements are required.At present,most of the soldiers’ physical fitness training assessments use manual assessment methods,which are costly,inefficient,and susceptible to subjective factors.Therefore,this paper applies the human skeletal key point detection technology to the evaluation of soldier physical fitness training,captures the soldier physical fitness training action video by camera,uses the skeletal key point detection technology to obtain the human skeletal key point information,then uses the video action detection model to fuse the human skeletal key point detection information,locates and classifies the action sequence of the captured physical fitness training video,and finally extracts the action key frames for Finally,the action key frames are extracted for analysis and evaluation,and a soldier physical fitness training evaluation system based on human skeletal key point detection is designed.The main research works in this paper are as follows.(1)The human skeletal key point detection algorithm based on pose flow compensation is proposed for the problems of low accuracy,large number of parameters and high complexity of the OpenPose network model to detect human skeletal key points in video streams.The algorithm uses the lightweight improved OpenPose model to obtain the skeletal key point information,then compensates the shortcomings of the OpenPose network model by constructing the human pose flow information between frames to improve the detection of video data,and finally optimizes the detected human skeletal key points to complete the key point coordinates that are defaulted due to severe occlusion.The algorithm proposed in this chapter achieves an accuracy of 95.2% on the MPII dataset,which improves the accuracy by8.0% and reduces the parameter amount by 71.5% compared with the original OpenPose,improving the recognition rate of human skeletal key points of video streams while reducing the network complexity.(2)A boundary-based local-global action detection algorithm is proposed for the problem of low accuracy of BMN boundary timing action localization.The algorithm selects the I3 D dual-stream fusion network as the feature extraction network,then uses the boundary-based local-global action detection algorithm to detect the temporal actions to obtain more accurate action sequences,and finally uses the S-OHEM strategy to optimize the sample ratio imbalance problem that occurs during the training process.Through experimental validation on the THUMOS14 dataset,the recall rate of temporal action nomination detection is improved by 2.3% compared with BMN,and the recognition rate reaches 98.8% on the KTH dataset and 99.2% on top of the self-built dataset.(3)A soldier physical training evaluation system was designed based on the above research.By acquiring the information of human skeletal key points during the physical training of soldiers and locating and classifying the action sequences in the training videos,and then comparing and analyzing the key frames in the action sequences with the standard actions,the automatic evaluation of the physical training actions of soldiers is realized.In the experimental test under the simulated soldier physical training scenario,the soldier physical training evaluation system designed in this paper has good practicality.