Research On Human Lower Limb Motion Recognition Based On Pressure Sensor And Accelerometer

Human motion recognition is an emerging research in the field of pattern recognition,which is of great significance to the development of intelligent interaction,medical health,bionic robot and other fields.With the development of microelectronic technology and wearable device,human motion recognition based on single acceleration sensor has become a popular research direction because of its portability,comfort and low cost.Though many related research results have emerged at home and abroad,its development still faces many difficulties and challenges.Aiming at the problem that it is difficult to distinguish the starting and ending points of a single action based on data stream of single acceleration sensor and achieve accurate identification of single action,this paper mainly takes a single acceleration sensor to identify 6 kinds of daily lower limb motion and falling action,supplement by thin-film-type resistive pressure sensor.The main work of this paper is as follows:(1)According to the needs of lower limb motion data acquisition,a set of human lower limb motion data acquisition system using wireless transmission which is wearable and able to monitor and store real-time data is designed.And some details on the hardware selection and software implementation are given.6 kinds of daily lower limb motion data from 14 volunteers and falling action data from 4 volunteers is collected by the system.(2)Aiming at the problem that it is difficult to distinguish the starting and ending points of the motion data in the acceleration data stream,a method of preprocessing of the acceleration data based on pressure is proposed,which can divide the starting and ending points of the action and intercept the acceleration fragments of the single action to achieve the recognition of the single action.In order to classify,FFT coefficients and DCT coefficients are firstly extracted.And different classification algorithms including KNN,SVM,Na?ve Bayes,and Random Forest are used to classify.Advantages and disadvantages are summed up after the comparison of the proposed method and the conventional pretreatment method using sliding window.Aiming at the problem of high rate of misclassification of partial action fragments for pressure-based acceleration data preprocessing,energy and standard deviation are added as feature and then 9973 action fragments received a maximum recognition rate of 97.01%.Finally,in order to simplify the features,LDA algorithm is used to reduce the redundant features.(3)Research on the fall detection method was specially carried out.According to the distribution of gravity in the coordinate system of the chip,a method of filtering out most of the normal short-term action and detecting the suspected fall action is proposed.In order to further determine the fall,5 layer wavelet decomposition of the suspected fall sequence base on ‘db4' wavelet is carried out.Three types of features including wavelet energy of the third and fourth layers,the angle between the gravity and the negative direction of z-axis and quartile deviation are extracted.SVM is used to classify and the accuracy of the fall action recognition reached 98.04%.Overall,the study of accelerator-related human motion recognition is one of the important research contents of pattern recognition and wearable computing,and there are still many applications and values that need to be excavated by researchers in the future.