Research On Human Complex Activity Recognition Method Based On Mobile Terminal

Human complex activity recognition is an important issue in wearable and mobile computing.Using mobile phones for human complex activity recognition is of great significance for observing the daily activity of users.At present,most researches focus on identifying simple activities(walking,running,sitting,etc.).Simple activities are characterized by repetitive actions or single posture.Existing methods based on complex activity recognition have insufficient extraction features,which leads to easy confusion between similar activity.In order to overcome the recognition errors caused by incomplete extraction of existing methods,this paper studies a hybrid network based on one-dimensional convolutional neural networks(CNN)and bidirectional long short-term memory(BLSTM)networks Human complex activity recognition method based on model and multi-feature fusion.Because the single-layer network output cannot obtain the best state information,this paper studies a method of human complex activity recognition based on data separation and multi-scale features,which can further improve the accuracy of complex activity recognition.Human activity is often closely related to the location of the occurrence.This paper studies a method for identifying complex human activity based on location coding,so that the network model can exclude activity events that are unlikely to occur at the current location based on location information.This article finally designs and implements a Human complex activity recognition system of mobile terminal.The main research contents of this article are as follows:1.Study a human complex activity recognition method based on CNN and BLSTM hybrid network model and multi-feature fusion.In this method,a CNN network model is designed to extract the spatial dimension features of the data,and the temporal features of the data are extracted using BLSTM.Considering that in the process of activity recognition,the state output at the current moment is not only related to the previous information,but also related to the subsequent information.Using a two-way long-term and short-term memory network,the temporal context of the state information can be used to improve the accuracy of activity recognition.The method also designs a feature selection method based on sequence forward selection and network weights,and performs feature selection on the features extracted by traditional methods to obtain the dominant feature set.In order to further improve the accuracy of activity recognition,the superior feature set is fused with the feature vectors extracted from the CNN and BLSTM hybrid models,and the sensor data features are fully mined to achieve high-precision complex human activity recognition.2.Investigate a method of human complex activity recognition based on data separation and multi-scale feature combination.Multi-scale features can extract deep and shallow features of the network to obtain global overall information and local detailed information.By combining different types of network features,feature vectors with different characteristics can be obtained.Based on the combination of multi-scale features,the sensor data of the wrist and lower limb positions are separated and independently trained to obtain their respective state encoding information,which can overcome the problem of incomplete expression ability of a single neural network,thereby improving the recognition of complex human activity.3.Research on a human complex activity recognition method that fuses position information,one-hot coding the position information where the complex activity occurs,overcomes the insufficient feature extraction of one-dimensional position data,and encodes the encoded position information with the network output The feature vector and the optimal feature vector of sequence forward selection and network weight selection are combined as the input feature vector of the next fully connected network.This method can eliminate activity events that are not related to location information and improve the accuracy of human complex activity recognition.4.Design and implement a human complex activity recognition system based on a mobile terminal,complete the real-time recognition of complex activity,and verify the feasibility and effectiveness of the method in practical applications.