Study Of Human Activity Recognition Based On Single Accelerometer

Human activity recognition is an emerging research subject with very important theoretical research sense and application prospect in the artificial intelligence and pattern recognition field. With the development of micro electro mechanical systems(MEMS), the human activity recognition based on accelerometer has been paid attention by more and more researchers. Compared with the approach based on computer vision, accelerometer is characterized by simple device, low cost, abundant information amount, high sensitivity, low energy consumption. Moreover, it is not subject to the limitation of time and space.The current research on human activity recognition based on single accelerometer has come through the phases from theory stage to practical application. However, there are still many problems need to be solved, such as how to scheme out a reasonable feature extraction method aimed at practical application, how to find a more effective method of feature selection, how to design an efficient classifier with high classification accuracy, low complexity and strong generalization ability. Therefore, the main research of this thesis can be concluded as follows:1) The existing activity recognition methods are summarized. Compared with the activity recognition based on computer vision, the superiority of the approach based on accelerometer is systematic analyzed, as well as its process and existing technologies.2) An efficient feature extraction method is studied in human activity recognition system based on accelerometer. Based on the distribution characteristics and the time-frequency analysis, 8 features according to the acceleration signal are extracted. To achieve the optimal feature subset, the original samples with high dimension are reduced by principle component analysis(PCA) to get the best expressive features, then, through linear discrimination analysis(LDA) to get the best classification features. 7 activities of daily living are identified by support vector machine(SVM). Results of tests show that the feature extraction algorithm combined with PCA and LDA can effectively raise the classification accuracy of various activities.3) To obtain a effective feature selection algorithm, a modified method is to use LDA based on genetic algorithm(GA). The eigenvectors of the between-class scatter matrix are chosen to do genetic operation as chromosome to reduce the training error. Compared with other widely used algorithms, the feature selection method used LDA based on GA can effectively reduce the feature dimension and decrease the training error. Besides, it can also raise the classification accuracy.4) Referring to improve the classification accuracy and generalization ability of human activity recognition system based on acceleration signal, a back propagation(BP) neural network classifier trained via the hierarchical genetic algorithm(HGA) is utilized. A three-layer chromosome hierarchical structure is used to optimize the structure and parameters of BP neural network simultaneously. A new fitness function is proposed, meanwhile, improved selection, crossover and mutation operator is beneficial to joint optimizing the complexity and accuracy of network. Results of tests show that the improved BP neural network classifier based on HGA can effectively control the network structure and parameters. The average accuracy rate of test data is 95.52%.