Research On Human Motion Capture And Recognition Technology Based On Wearable Sensors

With the development of technologies such as micro-electromechanical,artificial intelligence,and the Internet of Things,the intelligence level of wearable devices has gradually increased.Human motion capture and recognition technology based on wearable sensors has gradually become a research hotspot in industry and academia.By analyzing the movement information of different parts of the human body collected by wearable measurement nodes,to capture,track and indentify human movements in real time or offline,it is possible to enable the external environment or equipment to intelligently perceive the p osition and orientation of the person,and understand and simulate the movement of the human body.It has important application prospects in smart manufacturing,medical rehabilitation,virtual reality,sports and other fields.This paper focuses on the technology of human motion capture and recognition based on wearable sensors.Following the principle of step-by-step,this research is carried out from four aspects:multi-sensor information fusion,human position estimation,human motion capture,and motion recognition,to improve the accuracy of human posture and position estimation and motion recognition and provide theoretical supports for applications in different fields.The main works of this paper are as follows.Aiming at the low orientation estimation accuracy of wearable measurement nodes,this paper proposes a single node orientation estimation algorithm based on multi-sensor data fusion.For different types of sensors,corresponding calibration methods are designed to improve the measurement accuracy and stability.Aiming at the problem that the magnetic interference coupling affects the calculation of roll and pitch angles,by analyzing the mechanism of magnetic field measurements in multi-sensor fusion for orientation estimation,this paper constructs a vector orthogonal to the gravity field instead of the magnetic field as an observation reference to correct the yaw estimation calculated by the gyroscope and avoid affecting the roll and pitch estimations.Experimental results show that the proposed single-node orientation estimation algorithm can achieve robust roll and pitch estimation under magnetic interference,and maintain higher orientation estimation accuracy than the comparison algorithms in a wide range of pitch motion and static and dynamic motion.Aiming at the error accumulation problem in human position calculation,a foot position estimation algorithm based on adaptive zero-velocity update is firstly proposed.Utilizing the staggered peak characteristics of the acceleration and angular velocity measurements collected by the foot-mounted measurement node,an adaptive simulated energy consumption curve is constructed as the zero-velocity detection feature to accurately identify the stance phase of human motion and correct the velocity integration error through zero-velocity update.The kinematic chain model of human joints is constructed.Based on this,the orientation and position of main joints are modeled to achieve the human motion capture.A position estimation strategy that fuses zero-velocity update with a kinematic model is formulated to improve human position estimation.Experimental results demonstrate that the proposed method is able to identify zero-velocity phases in multiple locomotions,achieving position estimation error less than 3%.Aiming at the problem of accurately recognizing human activities when labeled data is scarce,a human activity recognition algorithm based on neural network and contrastive learning is proposed.A contrastive learning feature encoder based on deep convolutional transformer is designed to extract local and hierarchical features while focusing on the dependence of global features on longer distances,and learn activity feature representations from unlabeled sensor data.A randomly combined data augmentation strategy for input data is designed through detailed experiment,to enhance data diversity and enhance the ability to learn discriminative features.The performance of the proposed model is evaluated on datasets in three application scenarios of daily life,medical monitoring,and intelligent manufacturing.Experimental results show that the proposed method can achieve the mean F1 scores of 95.64%,98.40%,and 88.39%on UCI-HAR,Mhealth and Skoda datasets,which outperform the compared six popular recognition algorithms.