Research On Human Posture Recognition And Trajectory Reconstruction Method Based On Inertial Sensors

Human posture recognition and trajectory reconstruction based on inertial sensors have important applications in the fields of rehabilitation medicine,sports training and physical games.Based on the inertial sensor signal and the theory of deep learning algorithm,this paper conducts an in-depth study on the human body’s daily motion posture recognition including sitting,standing,upstairs,downstairs,walking,running and cycling.Based on inertial measurement theory,the short-time trajectory of limbs was reconstructed.The main work completed is as follows:(1)Designed and implemented the inertial sensor node hardware circuit.Combining the actual application requirements of miniaturization and low power consumption of data acquisition devices,the data acquisition circuit is designed and implemented with STM32F103C8T6 as the microprocessor and MPU6050 as the inertial sensor.(2)Inertial sensor random error modeling and correction.The random error of MPU6050 inertial sensor is identified by Allan’s variance analysis method.After data pre-processing such as constant component separation and trend term removal,the random error of MPU6050 is modeled by ARMA model.On this basis,the Kalman filtering method is used to correct the random errors of MPU6050 data.The experimental results show that the random error of the inertial sensor after Kalman filtering correction has been significantly improved.(3)The Inception-LSTM posture recognition algorithm,which introduces the channel attention mechanism,was designed to address the limitation that the traditional machine learning-based human posture recognition requires manual extraction of data features.The algorithm consists of the Inception convolutional structure to automatically extract the spatial features of the data and the LSTM network to extract the temporal features of the data.On this basis,the improved ECA module of the channel attention mechanism is introduced into the neural network classification model to focus the model on the key information of human posture.The experimental results on the publicly available dataset DSAD show that the average recognition accuracy of the proposed model is 98.10%.The analysis of the visualization results of channel attention weights shows that the recognition of human posture by the proposed model is interpretable and consistent with the intuition of daily life.(4)An attitude solving algorithm based on gradient descent and complementary filtering was designed.The algorithm consists of gradient descent for iterative optimization of attitude quaternions and complementary filtering for fusion of attitude angles.On this basis,a limb action localization method based on threshold switching is proposed by combining the characteristics of short-time limb motion,and the variance threshold is set by using the characteristics of the variance of the combined acceleration at the beginning and end of limb motion to accurately capture the starting and ending positions of the action.The experimental results show that the drift of pitch angle based on gradient descent and complementary filtering solution is about 0.4° and the drift of transverse roll angle is about 0.2° in the stationary state of about 10 s.In the dynamic experimental test,the algorithm can effectively improve the stability of attitude angle solution.The motion capture method based on threshold switch can effectively locate the starting and ending points of the motion and finally reconstruct the shorttime motion trajectory of the limb.(5)Comprehensive experiments and performance analysis of the system were completed.The results of the system comprehensive experiments show that the accuracy of the proposed pose recognition algorithm reaches 97.98% on the self-built dataset,which is 3.04% and 5.66%higher than the accuracy based on CNN and LSTM networks,respectively.The proposed limb trajectory reconstruction method can effectively suppress the divergence of the Euler angles and finally reconstruct the trajectory of the aerial writing digital action in the wrist part better.The posture recognition and trajectory reconstruction algorithm proposed in this paper can better describe human motion,laying the foundation for further motion analysis and providing some guidance for future related research.