The Human Activity Recognition Based On Sensor Data

Since 1980 s, human activity recognition field began to develop, attracting a lot of researchers into this area. With the development of artificial intelligence, machine learning and intelligent hardware and MEMS technology, mobile devices now have a variety of sensors and powerful computing ability. With these resources, mobile device can collect abundant human activities data, using for further analysis. Especially in the personal health, transportation, navigation, commercial advertising and other fields, the demand for perceived user activities becomes more and more large.In this thesis, we mainly aim to detect and recognize the activities about the human body, by analyzing and processing the original sensor data generated from the human movement.We analyze and abstract the action of human body, find out the feature of the action, and then construct the corresponding action model. At the same time, we construct the human action model for a kind of movement based on sensing data, and propose a detection algorithm which is used to divide and extract the action segment accurately. Finally, we obtain the feature vector from the motion segment data, and use the classifier to train and classify the detail kinds of movements. In this paper, we mainly focus on the kind of actions about the lifting arm movements, including the action of smoking, drinking water and scratching head.The core of the research mainly focuse on two aspects: the construction of action model and extraction of the action segments, as well as detail action classification recognition. In the first part, we presents an algorithm to extract the action segment from the original sensor data viaing the abstract analysis of the lifting arm movement. At the same time, it is able to extract a single action segment accurately from the data set, and accurately detect the change points of each state in a single segment makes it possible to automaticly obtain the segments of the motion in the experiment and reduce the human involvement. It makes the machine understand and recognize the action. In the second part, based on the cross validation of the training sample set and the classifier model constructed, we successfully detect and recognize the detail actions in precision tolerance range. It verifys the whole recognition scheme proposed in this thesis that it's feasible and effective.Finaly, this thesis implements an experimental system. We use the data sets collected about 500 actions from 5 volunteers. The classifier is trained and constructed by using the training sample set of the feature of actions, and the result of activities recognition is obtained. We test the system in actual environment, the experimental result show that the accuracy of the recognition rate of the above three kinds of movements can be achieve 83.27%.