A Research On IoT Posture Recognition Technology Based On Multi-source Information Fusion

Posture recognition based on wearable sensor devices is the mainstream way to judge posture categories,and it is also a research hotspot in the field of wireless sensor networks and artificial intelligence.In the traditional research of human pose recognition,the problems of single motion signal acquisition and low recognition accuracy lead to inconvenience in carrying,high energy consumption and long-term uninterrupted use by increasing the types and number of sensors.Based on this,this paper studies the attitude recognition method of the Internet of Things based on multi-source information fusion,aiming at extracting the most representational motion feature information and achieving high-precision attitude recognition effect.Aiming at the application scenario of long-term high-frequency acquisition and portable convenience,a attitude recognition system based on LoRa wireless network with low energy consumption,long-distance transmission and high accuracy is designed and implemented.Specifically,the main research results are as follows:(1)To solve the problem of human motion information acquisition,the the acquisition and processing method based on multi-source information fusion is adopted.Specifically,the acquisition motion information of a single placement multi-sensor is selected,and compared with the mainstream multi-placement scheme,a plurality of sensors can fully acquire information to improve recognition accuracy,and a single placement position can satisfy the convenience of carrying.This paper first introduces the acceleration,gyroscope,magnetometer and single chip microcomputer used in the scheme.Then the original data from sensors are processed,and a recursive averaging filtering method is proposed to denoise the signals to solve the problem of noise.Sequential signals can not well represent the attitude characteristics,so the sliding window method is used to segment the data stream,and then the sensor data from a single window are extracted in time domain and frequency domain.Then,according to the data requirements of subsequent classifiers,the feature orientation is obtained.quantitative standardization.Finally,according to the above steps,data set samples of walking,running,jumping and going upstairs based on three kinds of sensor information are collected for subsequent use.(2)Aiming at the problem of too large feature dimension,a feature selection algorithm based on Filter-Wrapper is proposed.Firstly,MRMR algorithm is used to rank 162 features according to the evaluation criteria of high correlation and low redundancy,and the top 21 features with high scores are selected to achieve feature screening.Afterwards,the search algorithm of SFS,SBS and the classification accuracy of classifier are used to select the package feature,and the final selection of key feature vectors is completed.The feature selection method can effectively calculate the representation ability of each feature vector,and the feature subset can be interpreted,and then the classification influence can be inferred to the sensor level.According to the contribution of various sensor information to the classifier,abandoning the sensor with low contribution can further reduce the energy consumption of wireless nodes and meet the original intention of low energy consumption of the system.(3)Aiming at the problem of human pose recognition,based on the research of mainstream recognition algorithm.In this paper,the attitude recognition model based on Stochastic Forest model is constructed.In order to verify the effectiveness of the classification model,several commonly used classification models,such as support vector machine and decision tree?are also constructed for experimental comparison.The experimental results show that the Random Forest algorithm is superior to other algorithms in classification and recognition accuracy of attitude recognition.Finally,in order to further verify the proposed posture recognition method,experimental verification is carried out on the self-collecting dataset and the classical dataset.Based on the self-collected dataset,the classification accuracy is 98.9%when the number of features is 4.Based on the classical dataset,when the number of features is 5,the classification accuracy is 96.3%,which satisfies expectations.(4)In view of the application background of long-term monitoring attitude data for specific individuals or groups,a attitude recognition system based on LoRa Internet of Things technology is designed based on the relevant conclusions of feature extraction,classifier category selection and specific parameters.It has the characteristics of low power consumption,low cost,high recognition accuracy and long-distance wireless transmission,and can meet the needs of large groups,frequent and long-term identification.The whole system can be booted directly without user training.The system is divided into four modules:sensor sensing,LoRa wireless transmission,server attitude recognition and front-end interface display.After the system is developed,it can run stably and has strong expansibility.