Research On Human Behavior Recognition Algorithm Based On Wearable Inertial Sensor

The advancement of wearable sensor technology has brought great convenience to human daily life.These fast-developing sensor technologies have shown strong advantages in sports assistance,medical care,and safety monitoring.In recent years,electronic technology and small sensor systems have developed extremely rapidly,especially now that almost everyone's essential smart products are integrated with micro sensor systems,making research based on inertial sensors of great application significance.At present,there are more and more researches on human behavior recognition technology.The development of human behavior recognition based on video technology is relatively early,but it will be affected by light or obstructions,and has the inertia of small size,low cost,and easy portability.The sensor highlights its superiority,making the human behavior recognition technology based on inertial sensors a hot research topic now.Therefore,combining human behavior recognition based on wearable inertial sensors with people's daily lives can effectively promote people's health management and abnormal behavior detection.Multiple sensors can collect information on human motion recognition relatively comprehensively,but they will increase the user's experience burden and reduce comfort.At present,the application of classifiers is relatively mature,and the technology of feature extraction and selection still needs in-depth exploration.How to select suitable features,how many dimensional feature values to select,and what feature selection method to use,can make the calculation more efficient.The complexity is low,and the accuracy of recognition is very important.Therefore,this thesis designs a single-sensor-based human behavior recognition algorithm,which mainly focuses on detailed research on feature extraction and selection,and experimentally verifies the proposed multi-feature fusion algorithm through two classification methods.The main content is divided into the following parts:1.Explain the research background and significance of this thesis,and introduce the current research status of human behavior recognition at home and abroad,and the current challenges based on inertial sensor technology.2.The overall framework of this thesis is designed,including four aspects: sensor wearing position,data preprocessing,feature extraction and selection,and classifier selection.3.A multi-feature fusion algorithm with rotation angle is proposed.The information of the rotation angle is helpful for recognizing the change of the action in three-dimensional space.Secondly,in view of the high feature dimension in the feature extraction process,this thesis designs and studies a combined dimensionality reduction method,which combines principal component analysis and feature subset selection algorithm to effectively reduce the feature dimension.The 57 feature dimensions extracted are first reduced to 21 dimensions through the principal component analysis algorithm,then the optimal feature subset is selected using the packaging method,and the dimension is reduced to 9.Finally,the two feature vector sets with rotation angle and only time domain and frequency domain features are respectively identified and classified.Experiments show that the accuracy of the multi-feature fusion algorithm proposed in this thesis is higher,and the overall improvement is 6.5%.4.Two classification algorithms are selected.Hidden Markov model can effectively verify and explain the multi-feature fusion algorithm,and the fuzzy c-means clustering algorithm is a new application in the recognition of sensor data.5.Build a self-collected and self-built data set to train and test the human behavior recognition model proposed in this article.The results verify that the classification and recognition accuracy rate of the hidden Markov model reaches 92.5%,and the recognition efficiency is significantly better than other classic classifiers.The classification method based on fuzzy C-means clustering in unsupervised learning,on the basis of ensuring the recognition rate,the execution efficiency is higher and the time is shorter,and the final recognition rate can reach 95.5%.Finally,a comparative experiment was carried out for the wearing position of the sensor.For the movement of the lower limbs,the recognition rate of a single sensor worn on the back is higher than that of worn on the right front hip.