Design And Implementation Of Human Activity Recognition System Based On Smart Phones

The widespread presence of motion sensors on users’ personal mobile devices has spawned a growing research interest in human activity recognition. Smartphones are becoming a powerful platform for event recognition due to the number of sensors they are equipped with. This provides an opportunity to apply machine learning techniques on movement data in order to recognize people’s locomotion activities without changing their routine. In the event of natural disasters, the rescue is difficult to cast in terms of post disaster specific degree and field data scarcity.If we can understand promptly to the active state of the trapped personnel, this will greatly promote the rescue work. Moreover, recognizing common activities could help people maintaining their energy balance by developing health measurement and intervention tools, investigating the relation between common diseases and levels of physical activity, and providing feedback to motivate individuals to exercise and keep fit.Most of the proposed studies benefit from off-line classification approaches,where the data are collected on the cell phone but trained and classified off-line on a a backend server. The offline learning process typically involves a batch algorithm,which usually requires several passes over the dataset until convergence to the optimal model. Since the process will take a certain amount of time, researchers often perform this procedure on the computer. In order to develop practical and advantageous applications, classification of activities should be performed on-board the mobile devices especially for health and well-being applications. The machine learning method of online classification can recognize activity immediately when the new data arrives, so that the real-time system becomes possible. This thesis aims at near real-time and convenience, and builds a mobile human activity recognition system.In this thesis, the process of human activity recognition on smart phones are firstly analyzed, and then a new learning method for online activity recognition on smart phones using the built-in accelerometers is proposed, called CluRF, which creates the cluster representation consisting of clusteroid and field size for each type of motion data and combines with the base classifier to complete the classification task. To learn the base classifier model parameters and the initial value of the cluster representation set, an open data set and the collected acceleration samples are first preprocessed. The system designed in this thesis uses this method, and has several attractive benefits: Due to the classification model with the default parameter values, the user can directly use the system without suffering the training phase; Owing to the cluster representation for each type of motion data, it can quickly identify current activity; In order to achieve a higher accuracy rate, the user can use the training mode for quick model adaptation. In order to implement this system, an application on the Android platform has been developed. Finally,the training and no training models of the system are tested, and from two aspects of the accuracy and execution time, the proposed CluRF algorithm and the other two classification algorithms were compared. Experimental results show that the system not only has the ability of accurate online classification, but also can quickly adapt to user activity profile changes.