| In the rapidly developing era of big data,human activity recognition based on wearable devices has played an important role in indoor positioning and navigation,intelligent software design,and elderly care.However,there are still some shortcomings in the current research on human activity recognition.On the one hand,the coverage of data sets is not extensive enough and the types of activities targeted are relatively small;and On the other hand,the recognition effect in existing wearable device products is not ideal,which is prone to misjudgment.In response to the above issues,thesis collects data information on human activity behavior through self-made wearable devices,expands the UCI-HAR dataset,and serves as input data for the algorithm model.Then,the input data is classified and identified using two different deep learning networks.The main research content of thesis is as follows:(1)To address the issue of incomplete human activity behavior types in existing data sets,self-made wearable device is used to collect data information on abnormal human activity behavior,expands the UCI-HAR public dataset,and performs data preprocessing and extracts feature variables on the expanded dataset to provide input data for classification and recognition of subsequent network models.(2)Aiming at the problem of insufficient recognition effect in existing wearable device products,a hybrid network model of CNN-Bi LSTM-Self attention based on dual channel mechanism is introduced.Firstly,a dual channel approach is used to extract features from different types of input data,and then a convolutional neural network is used to extract spatial features of the input data.Finally,Long Short-Term Memory networks is used to extract temporal features of the input data.Subsequently,a self attention mechanism is introduced to further enhance the extraction ability of temporal features,thereby achieving high-performance classification and recognition of human activity behavior.Experiments show that the recognition accuracy of the CNN-Bi LSTM-Self attention network model reaches 99%,and the F1 score reaches 99%.(3)In order to solve the problem of long training time of the CNN-Bi LSTM-Self attention network model,a hybrid network model of Capsule GRU-Self attention is improved.The spatial features of input data are extracted through capsule blocks,and the temporal features of input data are extracted by superposing a self attention mechanism on the basis of gated neural units,thereby achieving classification and recognition of human activity behavior.Experiments show that the recognition accuracy of the Capsule GRU-Self attention hybrid network model reaches 98.4%,and the F1 score reaches 98.4%.(4)Design a human activity recognition and analysis system,expand the dataset for visual analysis,and process the algorithm model files used to identify and classify human activity behavior. |
