Research And Implementation Of Human Motion Detection And Recognition System With MIMU For Energy Harvesting

Energy harvesting from human motion is an emerging field of research, how to make the most use of the negative work done by the muscle to generate electricity is one of the present research hotspots in biomechanical energy harvesting. The currently available energy harvesting device can not judge the work status of the muscle, and lots of energies generated from the negative work during the motion are dissipated, so the energy harvesting efficiency needs to be improved. In order to improve the efficiency of the energy harvesting device with minimal body burden, both the movement posture of limbs and the motion pattern, acting as the standard to control the working phase of the energy harvesting device, should be acquired in real time.Based on theories of inertial measurement and pattern classification, key technologies of human motion posture detection and pattern recognition, for the energy harvesting device, are deeply investigated in this paper, the MIMU-based human motion detection and recognition system for energy harvesting is designed and implemented. This system can detect the posture of limbs and recognize the pattern of human motion in real time, and this research can lay the foundation for enhancing the power harvesting capability of an energy harvesting device.The main research works of the dissertation are as follows:(1) The hardware system is developed with STM32F103TB as the microprocessor and MPU9250 as the 9-axis MIMU. The main systematic errors of MIMU are analyzed, and the error model containing the drift error, the sensitivity error, the non-orthogonal error and the non-alignment error is established. Both the calibration method for magnetic sensor based on ellipsoid fitting and the global calibration method for MIMU are researched and implemented. This calibration research lay the foundation for improving the detection and recognition precision of the system.(2) The correction algorithm of angle degeneration is combined with the attitude calculation algorithm of quaternion based on the extended Kalman filter, and the stability of the attitude angle is improved. The auto-adjusting algorithm of the measurement noise covariance is designed for reducing the impact of vibration interference caused by human motion. The attitude calculation algorithm of quaternion based on the EKF has an important shortcoming that the yaw angle, the pitch angle and the roll angle would be affected if the magnetic field is distorted, so the attitude calculation algorithm of direction cosine based on the KF is designed. The attitude calculation algorithm of direction cosine based on the KF is used in combination with the attitude calculation algorithm of quaternion based on the EKF, ensuring that the yaw angle and the roll angle would be accurate even though the magnetic field is distorted.(3) The data partitioning algorithm based on double sliding windows is designed, and the time domain feature parameters of the standard deviation and the skewness as well as the frequency domain feature parameters of the discrete cosine transform are selected as the feature information. The advantages and disadvantages of the BP neural net algorithm and the SVM algorithm are contrasted by the simulation analyses. On this basis, training procedures of two human motion pattern classifiers based on BP neural net and SVM are completed, these two classifiers are used to recognize four motion patterns of walking, walk upstairs, walk downstair and running, and their classification effects are analysed comparatively; In terms of the false classification occured at the point of pattern switch predominantly, ameliorative measures are proposed, and the average correct recognition rates of the classifier based on BP neural net and the classifier based on SVM are improved.(4) Effects of the system calibration, the shank gesture detection and the motion pattern classification are experimentally validated. The validation result of the system calibration has shown that the yaw error reduced to 7.6 percent, the pitch error reduced to 22.7 percent, and the roll error reduced to 44.4percent. The validation result of the shank gesture detection has shown that the improved attitude calculation algorithm can detect the posture of limbs accurately. The validation result of the motion pattern classification has shown that the improved classifier based on SVM has a better effect on recognition, and the average correct recognition rate reaches up to 96%.In this paper, the function of detecting the posture of lower limbs and classifying four motion patterns of walking, walk upstairs, walk downstair and running in real time is implemented, providing a feasible solution for improving the efficiency of the energy harvesting device. The results of this paper have a certain reference value.