Research And Implementation Of Action Training System Based On Key Point Detection

Action training needs professional guidance,and non-standard or wrong body movements will have a bad impact on the movement effect.The traditional exercise training methods are mainly as follows:paper graphic learning method,classroom direct learning method,video recording learning method,etc.Although these methods are simple,they have many shortcomings,such as low learning efficiency,high learning cost,and unable to get feedback in time and so on.With the rapid development of deep convolutional neural networks,the key point detection technology of human bones based on deep learning has been widely used in human-computer interaction,intelligent monitoring,motion analysis and other fields.Combining the key point detection of human bones with action training can promote the research in the field of key point detection of human bones.And it can guide the training actions more scientifically and accurately,and promote the development of intelligent action teaching.This thesis first conducts a detailed investigation and understanding of the current situation in the field of human bone key point detection and motion training at home and abroad.Expounded the two-dimensional human motion features in the field of human bone key point detection,the classification of detection algorithms,and common dataset and corresponding evaluation index.Then,introduced and analyzed the top-down human skeleton key point detection network HRNet jointly proposed by the Chinese Academy of Sciences and Microsoft Asia Research Institute in 2019.In view of its complex network structure and large parameter volume,a lightweight high-resolution human bone key point detection was proposed.The network HS-GattNet was trained on the MS COCO 2017 dataset and compared with some existing SOTA methods.The experimental results show that the detection network can reduce the calculation and parameter of the model under the premise of ensuring the detection accuracy.Generally speaking,there are some differences in the speed of the same action between different environments or different people,so it is difficult to evaluate the similarity of two independent poses.Therefore,this thesis proposes a key frame based and segmented dynamic time warping algorithm to solve the problem of action misalignment.Firstly,the key frame extraction technology is used to remove the redundant frames in the training video.Then,according to the number relationship between the key frames in the training video and the template video frames,combined with the proposed key point detection algorithm,the dynamic time warping technology is used to align the human feature vectors in time sequence.The experimental results show that this method can not only guarantee the accuracy of action alignment but also improve the efficiency of action alignment.Finally,this thesis designs and implements the action training system for the traditional martial arts plum blossom boxing.The system can collect the action video,and the key actions in the training video are automatically extracted and compared with the action information in the standard template action library to give the improvement suggestions.To a certain extent,it can be used as an auxiliary training tool for plum blossom boxing learners to a certain extent.