Research On Activity Recognition Method Based On Placement Independence

With the rapid development of microelectronics technology,smart wearable devices are emerging.Human body activity recognition based on wearable devices is an important field of pervasive computing,which has attracted more and more scholars' attention.Human activity recognition has been widely used in medical,military,sports and other fields,but few studies have solved the problem of locationindependent activity recognition.Currently,sensor-based activity recognition problems generally assume that the placements of sensors are fixed.However,this condition is obviously difficult to meet in practical applications.Aiming at the problem that the accuracy of the traditional human activity recognition method based on triaxial acceleration sensor depends too much on the placement and mode of the acceleration sensor device,this paper proposes a neural network based on convolutional neural network(CNN)and long-short term memory(LSTM)to solve it.This activity recognition method can effectively improve the accuracy of human activity recognition with the placement and mode of acceleration sensor device being in non-fixed situation.First,preprocessing is performed on the acceleration data.Then,the preprocessed data is input to a triaxial acceleration self-adaptive structure based on LSTM and fully connected layer structure,and this self-adaptive structure can adjust different input data to a similar coordinate system by itself.After this,the output of the structure is input to a CNN network model.Finally,the output is processed through a LSTM network,and the output of LSTM is into softmax and get final classification result.The network model combines the characteristics that CNN can adaptively extract deep features from shallow features,with the advantages of LSTM network to extract the context information of temporal domain feature.Therefore,the results of this method's performing deep feature extraction and classification are more reliable.This paper builds and evaluates the performance of the proposed model using the Daily and Sports Activities Data Set public dataset and the datasets collected by ourselves for location-independent issues.The experimental results show that the depth model proposed in this paper can effectively improve the overall activity recognition performance,and the accuracy rate on public and non-public data sets can reach 82% and 87% respectively.At the same time,the method of this paper is compared with the traditional activity recognition method to verify the effectiveness and advancement of the method.