The Human Body Posture Recognition Sysytem Based On Multi-sensor

The human-machine interactive system is a system where information is exchanged,understood and feed backed between man,machine and environment.How to more effectively implement the human-machine communication and understand human has become the main research focus and direction of future man-machine interaction,and human body gesture recognition technology as its core technology has also been paid more and more attention.At present,the mainstream research direction has two parts: one is the technology to focus on the human body motion capture and virtual reality;and the other is the human body gesture recognition technology to focus on the daily application and health care,whose purpose is to understand the typical behavior of the human body so as to complete the man-machine interaction.According to the data acquisition modes,the recognition technology is divided into two categories,namely,the recognition technology based on image and the recognition technology based on sensor.The recognition based on image is non contact-type,and applies to the man body posture recognition in the fixed area while the recognition based on the sensor can follow the target for data acquisition,having no geographic restrictions and on the basis of current wearable technology,the sensor will not affect the contact point node movement.Hence,this paper will design the human bod y posture recognition system based on sensor used in daliy situations.The recognition system includes "walking","running","strenuous exercise in situ","stand","sit","lie","falling" and "normal transition state".First of all,this paper designs a portable data acquisition terminal.Based on the basic requirements of data acquisition and portability,the terminal will be divided into four modules: the sensor module,control module,memory module and power management module.Terminal is stable in data acquisition with a long standby time,and is the original data source for human body posture recognition system.Secondly,in order to obtain relatively pure data,this study will make a further pretreatment of the raw data and feature extraction and data.Pretreatment includes low-pass filter noise,scaling,and removal the disturbance of acceleration of gravity.Feature extraction includes time domain feature extraction and frequency domain feature extraction.In details,they are the mean value,standar d deviation of sample frame,the median absolute deviation,interquartile range,maximum,minimum,mean value of quadratic sum,AR model coefficients with berg order of 4,the serial correlation of all axes,information entropy feature,peak frequency poin t,secondary peak frequency point,bandwidth of peak,the weighted average frequency point,prominent frequency point quantity characteristic.After the comparison between different samples,the availability of characteristics can be evaluated.Finally,a human body posture recognition algorithm based on the characteristics mentioned above is built.First,"walking","running" and "strenuous exercise in situ" are defined as the motion state;the "stand","sit" and "lie" are defined as the static state;and the "falling" and "normal transition state" are defined as the transition state.Each state is to implement classification and recognition using different methods based on the characteristics of the categories such as support vector machine and decision tree.The finally build models constitutes a relatively mature human body posture recognition algorithm architecture.