Research On Human Action Recognition And Application Based On Wearable Sensors

In recent years,with the rapid development of information technology and artificial intelligence technology,human motion recognition has been widely used in daily life and work production.In the field of human motion recognition research,the way of collecting motion data through video or photography is easily affected by factors such as perspective and light,which is not easy to describe and extract features.To solve such problems,this thesis adopted human skeletal joint data obtained by sensors as the research object,and studied human motion recognition and application by combining CNN,LSTM and GRU models.The research work of this thesis mainly included the following aspects:(1)Constructed and processed human action datasets.The MSRC-12 dataset was analyzed and the homemade dataset Dataset-A was constructed.In the process of using wearable sensor equipment to construct DATASET-A,combining Euler Angle,rotation matrix and coordinate transformation,the three-dimensional coordinate calculation formula of human bone joint was summarized.In addition,in the process of processing node data,in order to improve the feature extraction efficiency of human movements,inspired by the idea of data image coding,12 data coding combination schemes combining3 data arrangement orders and 4 encoding methods were proposed for gray image,and 4encoding methods were adopted for RGB image.A total of 16 data coding combination schemes were proposed,and the node data was encoded as PNG image storage.(2)Built human action recognition model based on data coding and machine learning.Firstly,the CNN model was used to compare and analyze 16 coding schemes,discussed the influence of data arrangement orders and encoding methods on action recognition,and selected the optimal coding scheme,i.e.gray Case2 Zhi.Secondly,time series models LSTM and GRU were integrated to extract the features of action data in time series,and the action recognition effect was compared with CNN.Then LSTM,GRU and CNN are combined respectively to compare the human action recognition effect.In the experiment,coded and uncoded action data were respectively input into the same model for training analysis,which verified the effectiveness of data coding for improving the accuracy of human action recognition.The experiments show that the "Zhi" font encoding in the data arrangement order of Case2 is easier to classify actions,and the CNN-GRU model performs better in human action recognition tasks than CNN,CNN-LSTM and other models.And the recognition accuracy of CNN-GRU model on MSRC-12 data set is more than 98%.(3)Applied the data coding scheme and CNN-GRU model to the ROS-based wheeled vehicle for human-computer interaction experiment.Firstly,the wearable sensor device was used to obtain the real-time human motion data,and it was intercepted and encoded.This part realized the human motion data acquisition and processing module.Secondly,the encoded action data was input into the CNN-GRU model trained by DATASET-A for action recognition,and the human action recognition module was realized.Finally,combined with the action acquisition,processing,recognition module and the topic publishing mechanism of ROS communication,the action recognition results(A1-A5)were used as instructions to control the wheeled vehicle to perform the specified movement.The experimental results show that the control actions identified by the CNN-GRU model can effectively interact with the wheeled vehicle.