Human Dehavior Recognition Based On IMU Sensor

In recent years,with the rapid development of artificial intelligence technology and deep learning,the application scope of both has been gradually expanded.Human behavior recognition because of its wide areas,the application of the comprehensive and high-profile,particular application is in a lot of on the new direction of social science and technology,human body recognition gradually become the mainstream research direction,along with the continuously go deep into cross disciplines,as well as the constant innovation of scientific research personnel from all walks of life,human behavior recognition has made breakthrough progress in many fields.In this thesis,IMU sensor equipment is used to collect behavioral data.The main content of the study is as follows: the mainstream behavior recognition and classification methods are mainly studied.Based on previous research experience,the effect of the fixed threshold method is not good,so an adaptive threshold method is designed.The sensor data of waist and leg were collected by IMU sensor,and the differences of characteristic values of four kinds of human outdoor movements(walking,jogging,jumping,falling)were analyzed.Four kinds of acceleration thresholds and two kinds of human posture inclination Angle were used to distinguish the four kinds of human outdoor movements.During the experiment,different types of people were selected for experimental analysis to ensure the correctness of the experiment.Finally,the average accuracy of the waist test reached 96.65%,and the average accuracy of the leg test reached 98.85%.Combined with the data set made public by IMU in the current field and the ability of neural network to extract features from time series data,the experiment of RNN and LSTM network models is carried out,and the comparison and analysis are made respectively.It has a Recurrent Neural Network(RNN)with gradient disappearance and gradient explosion,so Long short-term Memory(LSTM)is used to optimize the experiment.Compared with the standard RNN and other deep neural network models,THE structure of LSTM itself is more complex,and its own gate control device can effectively improve the memory effect of long data.Compared with the experimental effect,LSTM is significantly better than RNN.In the experiment of LSTM,we made a comparison of adding attention mechanism.In the experiment of LSTM,the hyperparameter orthogonal experiment was carried out.In this experiment,the optimal combination of orthogonal parameters was obtained,and the recognition rate reached 98.70%.The experimental effect of optimized parameters was verified on other data sets.