Human Motion Recognition Method Based On Fusion Of Multi-Source Body Sensor Information

Body sensor networks based motion monitoring technology can provide users with effective health protection and treatment measures,monitor daily exercise of the elderly,quantify exercise levels and also provide users with timely physiological feedback.Body sensor networks based motion recognition technology is a new research field in recent years,and it belongs to the category of human action recognition.Sensors are mainly used to collect physiological signals of human body,and then recognize different activities with the help of pattern recognition method.Based on the motion data collected by a number of wearable inertial sensor nodes,some problems existed in motion recognition are studied,the main contributions of this paper are as follows.(1)A new feature selection algorithm-modified linear discriminant analysis(MLDA)algorithm is proposed.The traditional LDA algorithm has error disturbances when solving the optimal projection matrix.MLDA feature selection algorithm proposed in this paper uses the congruent transformation in matrix to solve the error disturbances when we estimate the within-class scatter matrix.Experimental result shows that the algorithm can improve the accuracy of human motion recognition.(2)A new classifier(KSVD-SRC)based on sparse representation algorithm(SRC)and KSVD dictionary learning algorithm is proposed.How to improve the accuracy of human motion recognition is worthy to make further studies.Aiming at the defects of complete dictionary in SRC algorithm,KSVD algorithm is used to update the original complete dictionary in KSVD-SRC classifier,which is good to get the relatively sparse coefficient vector.Experimental results show that the proposed method is much better than the traditional recognition algorithms.(3)A multi-classifier multi-sensor hierarchical fusion algorithm is proposed.In practical applications,just using a single-classifier to recognize some motions may not get good results.In this paper,a hierarchical fusion model is proposed,which includes the fusion layer of the classifier and the fusion layer of the sensor.The decision weights of each layer are mainly obtained by entropy method.The fusion model can effectively improve the recognition accuracy and robustness of the recognition system.(4)A framework for segmentation and recognition of motion data streams based on multiple inertial sensors is proposed.The data collected by inertial sensors is in the form of data flow.If we want to recognize the motions,we need to process the data stream.In this thesis the singular value decomposition technique is used to preprocess the motion data stream,and then the proposed similarity measure function-MSHsim is used to finish the fine segmentation of the motion data.Finally,the hidden Markov model is used to recognize the processed data.The algorithm framework has a lower computational complexity,more importantly,because the characteristics of the sensor data are considered in the algorithm,so it can effectively improve the accuracies of data segmentation and recognition.