Sitting Posture Recognition Based On Flexible Sensor

With the progress and development of science and technology and the popularization of information office,people spend more and more time sitting and working.The increase of sedentary time will lead to various diseases related to sitting posture.These diseases are often difficult to recover in a short time,which seriously affect the life and work of patients.Therefore,an external device that can detect the sitting posture state is needed to get a reminder in case of wrong sitting posture.In this paper,a sensing cushion based on flexible sensing material is made,which is combined with depth learning algorithm to achieve the purpose of sitting posture recognition.This paper focuses on the research of sitting posture recognition technology based on flexible sensor,and carries out four aspects of work:(1)Based on the knitted fabric piezoresistive flexible sensing material,three different structural sensing units are designed,which are single layer(SL),double layer(DL)and triple layer(TL).The chi650 electrochemical workstation was used to compare the three sensing units.In the pressure range of 0-120 kpa,the relative current variation of TL structure is significantly greater than that of DL and SL structure in each pressure section,which is easier to collect and distinguish data.Therefore,TL structure is finally used as the sensing unit to produce a total of 1024 sensing points with a size of 50 cm × 50 cm sensing cushion.(2)Using the secondary scanning method,the array resistance at the same time is collected twice by switching the direction of VCC and GND,and the coupling resistance is eliminated from the simultaneous equations of the data collected twice,which solves the cross coupling problem of array resistance caused by the conventional array acquisition circuit,The STM32F103VET6 microcontroller is used as the main control chip to realize the hardware acquisition circuit of the secondary scanning method.(3)The bad sitting posture was analyzed.The sitting posture was divided into five categories: upright sitting,forward leaning,backward leaning,left leaning and right leaning.Taking the conventional acquisition circuit and secondary scanning circuit as groups,two different data sets were collected through volunteers,and a three-layer BP neural network was constructed to verify the availability,advantages and disadvantages of the two data sets.(4)Based on the VGG11 network structure,the number of convolution cores in each layer is reduced to reduce the occurrence of over fitting,the two lower sampling layers are removed to increase the dimension of the feature map,and the full connection layer is replaced with the full convolution layer to retain the spatial location information of the feature.An mVGG-FCN network more suitable for the data set structure of this paper is constructed,which is combined with Le Net,mVGG-FC,mVGG-GAP Five kinds of BP neural networks are compared.The results show that the mVGG-FCN structure has the highest global accuracy of 98.46%in 50 epochs,and the recognition rate of each sitting posture category has reached more than96.4%.At the same time,it has the shortest running time of 3.81 s in the three network structures with VGG11 as the template.The experimental results show that the convolutional neural network based on mVGG-FCN structure is suitable for the application scene of sitting posture recognition.