Research On Motion Recognition Algorithm Based On MEMS Sensor

Motion capture is a technique that uses the corresponding sensor to sample the motion data of the target object,record its motion process,and then process the data through a computer to finally realize the function of classification,recognition and analysis of the motion of the object,including MEMS inertial measurement.The unit's motion capture system can truly realize motion data acquisition that is not constrained by the scene,and is the most promising in commercial development.At present,the research on motion recognition model mainly stays at the simple application level of traditional machine learning algorithms.The recognition algorithm is often only applicable to specific motion data sets,and the model's migration ability is poor.In order to make up for the lack of research status,this paper proposes to include The generalized action recognition chain process of "sensor signal reception-data preprocessing-feature engineering-recognition model",the main research contents are as follows:A standardized inertial motion data pre-processing flow is implemented,including a simple error calibration scheme and a suitable data cleaning method,which avoids the measurement error accumulation of the MEMS device and the influence of dirty data generated during the acquisition process on the data quality.On the basis of the data preprocessing,the action interception algorithm implementation under the two segmentation logics of event window and action window is introduced for the requirements of offline recognition and online recognition.In order to accurately detect the starting point and the end point of the motion,the Teager operator and Gaussian smoothing filter are used as the motion amplitude index,and the parameter modeling method of the action threshold is derived.The adaptive threshold determination scheme based on the energy peak is proposed,which can be more accurate.The ground intercepts various effective action signal segments.In order to cover the main inertial motion characteristics as much as possible,this paper presents a feature calculation method that can express multiple types of motion characteristics,including statistical features,signal time-frequency features and system modeling features.In order to avoid the overlap of the actual distinction between features,this paper proposes a set of practical feature contribution evaluation indicators based on the information gain principle.The data set can be used to filter and adjust the feature combination scheme.In the recognition model part,this paper discusses the shortcomings of the traditional machine learning algorithm as a motion recognition model.Based on the integrated learning principle,the classification regression model is used as the basis classification model,combined with the boosting strategy,and the extreme gradientlifting tree is implemented on the motion dataset.The learning algorithm,combined with the feature extraction strategy,can achieve an accuracy of 97.99% for 22 types of actions in the badminton basic motion data set.Under the online motion recognition scheme,because the sliding window cannot accurately locate the position of the motion data,certain requirements are imposed on the data displacement invariance of the recognition algorithm.Under this premise,a deep learning scheme of alternative feature engineering scheme is proposed.The one-dimensional convolutional neural network is used to construct the motion recognition model.The translational extended dataset under the sliding window acquisition scheme can achieve an accuracy of 96.83%,showing relative Robustness of data displacement.