Research On Human Daily Action Segmentation And Pattern Clustering

Human action recognition is an important research direction in the field of pattern recognition.In recent years,with the continuous development of microelectronics and wearable device technology,action recognition based on inertial sensors is more suitable for data collection of complex human movements due to its advantages such as portability,low cost,and protection of user privacy which plays an important role in medical,recreational,security and military fields.In the future,when mobile communication technology enters the 5G era,it will be more convenient for people to build micro sensor networks and apply them to our daily lives.In this paper,the dual inertial sensors are used as the main part,and the membrane pressure sensors are used as the auxiliary part to identify the daily movement of human body.In addition,the basic motion state of standing,squat,lying-down,sitting,walking and other daily actions of human body is analyzed and studied.The main procedures are as follows:1.According to the laboratory's existing human movement data collection system and the database collected and produced by the platform,it is used for data analysis and movement classification.2.Analyze and define the daily movements of the human body,and abstract the human body's approximate multi-rigid body kinematics model.Pre-processing and feature extraction of the original data in the database,pre-processing is to smooth the data filtering process,to eliminate the impact of high-frequency signals in the data on low-frequency action signals.The feature extraction is to extract relevant features in two aspects of statistical features and physical features and describe the daily movements of the human body in detail.3.Aiming at the problem that the inertial sensor cannot directly divide the start and end of motion,a data processing method based on dynamic segmentation of human motion based on sliding window is proposed,and human motion is divided into three basic motion states.4.Aiming at the data segment based on clear segmentation,the decision tree algorithm is used to judge the basic state of the action,and three corresponding neural network models are trained according to the three different states to perform actionclassification and recognition.Action features are often judged empirically when they are initially selected,and the decision tree algorithm recognizes features that are effective for classification and recognition,thereby reducing the number of features.By using the existing database to verify the relevant algorithms,it is shown that the rate of decision tree recognition the three states is above 95%,and the rate of the neural network's average recognition all actions is 96%.The experiment then verifies the effectiveness of the algorithm.