Study On Human Activity Recognition Based On Smart Device

With the progress of the age and the continuous development of science and technology, we can see the intelligent devices everywhere, such as Smart Band of XIAOMI, Google Intelligent Glasses, etc. Now most of the intelligent devices can collect sensing data through various built-in sensors and can realize the context-awareness service, health monitoring and life recording through rigorous analysis. Nowadays, human activity recognition based on classification algorithms has been widely used to identify the behavior category. Due to the low cost, and the recognition process is undisturbed for normal daily life, human activity recognition has been used in health monitoring, smart home services, sports injury prediction and other important areas. The availability and accuracy of human activity recognition system are two important factors which affect the user experience. So, what features of human behaviors need to be extracted are representative, how to identify the activities by these features, and how to minimize the energy consumption greatly under the premise of maintaining higher recognition accuracy are very important. According to the human activity recognition based on the smart mobile device, this thesis mainly made the following contributions:(1)The first part of this thesis proposes an implicit identity authentication system based on keystroke behaviors, and it is the first attempt to consider the changes of a user’s gesture. Built-in smart phone sensors are utilized to collect five keystroke features (i.e., keystroke acceleration, pressure, size, time, and device orientation) in the background when a user unlock the screen and match with the corresponding pre-established gesture models in order to determine whether the user is a valid user or not. This authentication method is more suitable for real life, as it can reduce the authentication system’s error rate by matching with the corresponding gesture models. This method and digital password are respectively regarded as implicit and explicit authentications for the identity authentication of smart phones. In other words, our approach can act as a second authentication method and make up for the inadequacy of digital password in a cost-effective and user-transparent manner.(2)The second part of this thesis presents the work efficiency monitoring system based on the smart watch. This system collects users’ behavioral characteristics in work.We conduct the data pre-processing and feature extraction for the data acquired, and use SVM algorithm to train the classifier and predict the test set category. It can determine whether the unknown state that the users work belongs to the serious state by the well trained classifier, and eventually provide the monitoring report to inform the situation of work efficiency to the user.(3)The third part of this thesis presents a method of activity recognition based on Energy-Efficient Schemes According to the basic idea of the first two works. In terms of data acquisition and processing, Energy-Efficient Schemes adopt the best sample rate and extract the most effective feature combinations in accordance with the different activities, so as not to increase the energy consumption; while in terms of recognition algorithm, the Energy-Efficient Schemes adopt the improved structure of multi-class SVM, combine it with the probability of activity occurrence, so as to reduce the time complexity of recognition. This method can minimize the energy consumption greatly under the premise of maintaining higher recognition accuracy.This thesis studies how to improve the availability and accuracy of human activity recognition system based on the above three parts.