Study Of Human Activity Recognition Based On3D Accelerometer

Human activities recognition based on accelerometer is an emerging research in the fieldof pattern recognition. The rapid development benefits from the continuous development ofmicroelectronics and sensor technology as well as the in-depth research of pattern recognitiontheory. With the growth of people demand for intelligent interaction and medical monitoring,the human activities recognition based on accelerometer receives a widespread attention in thearea of health care, motion detection and energy consumption. Compared with the activitiesrecognition based on computer vision, the approach based on accelerometer deals more withthe essence of human motion. Besides, it is not subject to the limitation of specific scenariosand time, and the less energy consumption and lower cost make it more suitable forpopularization and application.Although activities recognition based on accelerometer has made a great progress inrecent years, many problems still remain unsolved, such as how to extract signal features withstrong representation abilities, how to detect falling in practical application, how to designactivity classify algorithm with high accuracy and generalization ability. In order to solve theabove problems, the current thesis focuses on the following issues:1) The existing activities recognition methods are summarized; the approach based oncomputer vision and that on accelerometer are compared, and the advantages of usingacceleration signals to recognize activities are systematic analyzed, as well as its process andtechnology.2) From the start point of the time-frequency analysis and the distribution characteristicsof acceleration signals, with the help of wavelet analysis and other methods, two novelfeatures about the wavelet energy based on angle and the slope of key points connection areproposed, which depicts the signal characteristics from different aspects. The recognition ratesof different features sets are compared using the independent-dataset test and cross-validatedtest, and the results shows the effectiveness of this two features.3) In the fall recognition, the common classifiers require a large number of trainingsamples, the training samples tend to be collected by the way of deliberate and repeatedfallings, but it is very inconvenient for users. In order to solve this problem, a fall detectionmethod based on hidden markov models(HMM) and body angle is put forward. The impact offall sample size on the recognition result is reduced by dealing with the fall detection as theproblem of deviation between samples and learnt model. In addition, the information beforeand after states of research object is effectively remained by the method of timing analysis,which is more accordant with the laws of physics.4) In the daily activities recognition, the hierarchy genetic algorithm(HGA) is used to optimize the construction and parameter of the radial basis function neural networks(RBFNN)classifier and enhance the generalization ability and recognition accuracy. A fitness function isdesigned to reduce the complexity of classifier and increase the accuracy rate. Theinterquartile range(IQR) is used to improve the crossover of parameter genes. With the help ofthe combination of two kinds mutation, the optimization efficiency is enhanced. Theexperimental results show that the RBFNN classifier trained by the improved HGA commitsfewer output errors and has simple structure. The recognition rate of7activities can reach to91.54%.5) An acceleration signal acquisition system is designed, and the acceleration data ofhuman activities can be collected successfully.