Evaluation System Of Human Motion Accuracy Based On Machine Learning

With the development of technology and people's desire for health,the research on the accuracy judgment of human movement using sensors is becoming more and more fierce.Early researches were basically carried out around self-designed wearable devices due to lack of corresponding industry standards and lack of common platform equipment.Researchers mainly focus on how to collect data stably in complex environments or how to improve sensors and how to improve model quality.At the same time,the corresponding products on the market,taking the Motion Sensing Game as an example,sensor designs,data collecting,and model training are all done by manufacturers.If the data generated by users can be made good use of,the performance of the product can be further improved..At the same time,the security of modern personal data is becoming more and more important,and privacy protection algorithms such as federal learning have emerged.Ensuring the safety of user data will make users more willing to participate in the process of improving product performance.With the development of smart phones and other smart devices,computing power has increased rapidly,and models can be calculated and run on smart devices such as smart phones.By calculating the model locally on a smartphone and uploading the sharing model instead of data,the model can be jointly trained without revealing personal data.However,there are still many problems in developing a usable process framework and using user-uploaded models.This article uses smart phones and other smart devices as a platform to conceive and initially implement a supervised shared data action accuracy judgment system.According to the characteristics of the system,the algorithm and operation framework,loss function and regular function are selected to solve the problem of sensor difference,upload data difference,and sensor position difference.The possible imbalanced data sets are discussed,and the ready-made imbalanced data set processing algorithms are discussed.A method of continuous training using complementary unbalanced data sets is proposed to solve the problem of unbalanced data sets.Aiming at the problem of difficult and expensive data collection for anthropometry,a solution for sharing data was proposed.Based on the idea of federated learning,a framework for sharing data training with exchange models among different data sources is designed.In the verification of the data collected through experiments,the accuracy of human movements using this system has been improved.