Research On Smartphone-based Activity Recognition And Biometric Identification Technology

With the widespread use of smartphones equipped with a rich set of sensors, smartphones have become a mobile sensing platform which combines communication, computation and sensing. By using smartphone's built-in sensors such as accelerometer and gyroscope, we can obtain abundant sensor data. How to use accelerometer and gyroscope perform human activity recognition and biometric identification are the hot and difficult problems in research fields of mobile sensing.This thesis carries out research on how to use inertial sensor data perform human activity recognition and biometric identification. By the analysis of related techniques,this thesis presents a feature extraction method based on unsupervised feature learning techniques, on this basis, and then by applying this feature extraction method to activity recognition and biometric identification to evaluate it. The main work and contributions of this thesis are summarized as follows:(1) In view of the drawbacks of traditional features, such as dependent on domain knowledge, could cause significant loss of information, combined with the inherent characteristics of the sensor data, this thesis presents a feature extraction method based on unsupervised feature learning techniques for the inertial sensors data. This method first learn feature mapping on each channel of sensor data, and then concatenate all channels together. By using unsupervised feature learning techniques, this method can automatically learn features from the data, thus avoiding some drawbacks of the hand-crafted features. The channel-wise way of using unsupervised feature learning techniques to process inertial sensor data, is proved a suitable method by theory and experiments.(2) In order to evaluate the proposed feature extraction method for activity recognition, we extracts a variety of features, including different way of unsupervised feature learning, time domain features and frequency domain features, and then evaluate them using three classification methods including C4.5, Naive Bayes and Support Vector Machine. The results show that our feature extraction method outperform other features on test accuracy.(3) Based on the research about the key technologies of smartphone-based biometric identification, we design an classification algorithm for identification in order to meet the reliability and availability requirements, and then evaluate this algorithm based on the proposed feature extraction method. The Experimental results show that by selecting suitable parameters in the algorithm, it can well guarantee the reliability and availability requirements of smartphone-based biometric identification.